research paper based on power system protection

• Reference Manager
• Simple TEXT file

People also looked at

Review article, a review of machine learning applications in power system protection and emergency control: opportunities, challenges, and future directions.

• Department of Electrical and Computer Engineering, University of Manitoba, Winnipeg, MB, Canada

Modern power systems, characterized by complex interconnected networks and renewable energy sources, necessitate innovative approaches for protection and control. Traditional protection schemes are often failing to harness the vast data generated by modern grid systems and are increasingly found inadequate and challenging for some applications. Recognizing the need to address these issues, this paper explores data-driven solutions, focusing on the potential of machine learning (ML) in power system protection and control. It presents a comprehensive review highlighting various applications which are challenging to address from conventional methods. Despite its promise, the integration of ML into power system protection introduces unique challenges. These challenges are examined in the paper, and suggestions are provided to overcome them. Furthermore, the paper identifies potential future research directions, reflecting the progressive trends in ML and its relevance to power system protection and control. This review thereby serves as an essential resource for practitioners and researchers working at the intersection of ML and power systems.

1 Introduction

Power systems, the backbone of modern civilization, have evolved from traditional generation and distribution models to complex interconnected networks that incorporate renewable energy sources and smart grid technologies. This evolution presents both exciting opportunities and significant challenges in terms of power system protection and control, calling for innovative approaches to ensure system stability, reliability, and resilience ( Hossain et al., 2018 ). Even though the existing traditional power system protection and control methods are robust and have been well-developed over the last century, they have been built upon mathematical models that may struggle with the uncertainties and nonlinearities inherent in the complexity of modern grid systems ( Karlsson and Hill, 1994 ; Makarov et al., 2011 ). Therefore, in this rapidly evolving landscape, traditional methods are becoming inadequate to handle the complexity of the system. In addition, these traditional systems often fail to capitalize on the rich data generated by the modern grid, which holds valuable insights into system operation and behavior ( Yu et al., 2015 ; Syed et al., 2021 ). On the other hand, there is an urgent need for efficient and near real-time algorithms to analyze and make better use of these available data.

The potential solution to this issue might be harnessing the capabilities of modern artificial intelligence (AI) techniques, utilizing their advanced generalization and predictive abilities to navigate the complexities of power system operations. Particularly, the enormous amounts of data generated in the power system can be processed using powerful tools present in machine learning (ML), which is a subset of AI ( Qiu et al., 2016 ). It has the capability to learn from data, adapt to new conditions, and continuously improve performance with experience ( Chellappa et al., 2021 ). In recent years, ML has emerged as a significant research area, reflecting a broader trend across various scientific disciplines ( Badrinarayanan et al., 2017 ; Cui et al., 2021 ; Mahadevkar et al., 2022 ). Figure 1A illustrates the annual growth in the publication of ML papers, as indexed in Scopus ( https://www.scopus.com ) over the last 20 years. To construct the graphs showing the trends, the database was searched using keywords related to machine learning and power systems, with consideration given to publications from the year 2000 onwards. The swift increase over the last 5 years stands as a testament to the field's rapid advancement and the widespread interest it has attracted.

Figure 1 . (A) Number of ML papers over time, based on data from the Scopus database.; (B) Number of power system solutions with ML technics over time.

In addition, the capabilities of data-driven approaches make it a potentially invaluable asset in the era of smart grids and renewable energy integration ( Cui et al., 2021 ). Figure 1B shows the annual growth in the number of publications that adapt ML techniques to power system solutions. The substantial increase in recent years is indicative of the industry's progressive incorporation of these modern techniques ( Ernst et al., 2004 ; Hadidi and Jeyasurya, 2009 ; Rudin et al., 2012 ; Alimi et al., 2020 ; Zhao et al., 2022 ).

Despite its promising potential, the integration of ML into power system protection and control is still in its early stages and is not without challenges ( Mahadevkar et al., 2022 ). The objective of this paper is to offer a comprehensive review of ML applications in the realm of power system protection and control. It provides an in-depth examination of the strengths, limitations, and potential of various techniques as applied to these domains. Additionally, the paper discusses the opportunities and challenges associated with integrating ML into protection applications and suggests future research directions, considering emerging trends in both the fields of ML and power system protection and control. The remainder of the paper is structured as follows: Section 2 introduces the basic concepts and techniques of ML; Section 3 offers some of the key performance requirements in power system protection and control; Section 4 describes potential opportunities; Section 5 delves into the bottlenecks in applying ML to power system protection; Section 6 explores future directions; and Section 7 presents the conclusions.

2 Machine learning: basic concepts and techniques

Machine learning is a field that is concerned with the development and study of algorithms that can automatically find solutions to problems using input examples or training data ( Pedro, 2012 ). It is a multidisciplinary field which consists of statistics, computer science, linear algebra, and optimization, to mention a few. The ability to learn from data makes ML algorithms primarily useful for addressing highly non-linear problems such as classification and function approximation where it is very challenging or even impossible to model the relation between input and output using traditional techniques ( Ray, 2019 ); some examples of these types of problems are image classification ( Gonzalez, 2007 ), text identification ( Lecun et al., 1998 ), Atari games ( Mnih et al., 2013 ) and board game solving ( Silver et al., 2017 ). The learning process is often classified into supervised learning, unsupervised learning, and reinforcement learning (RL).

2.1 Supervised learning

Supervised learning is the ML task of learning a function that maps an input to an output based on example input-output pairs ( Simeone, 2018 ). It infers a function from labeled training data consisting of a set of training examples. Examples of supervised learning algorithms include Artificial Neural Networks (ANNs), Support Vector Machines (SVM), Decision Trees ( Safavian and Landgrebe, 1991 ), and Random Forests ( Jin et al., 2020 ). SVMs are used for both regression and classification tasks, using a technique that minimizes the error rate while maximizing the margin of decision. Decision Trees and Random Forests are often used in classification problems, creating a model that predicts the value of a target variable by learning simple decision rules inferred from the data features.

2.2 Unsupervised learning

Unsupervised learning, on the other hand, involves the use of ML algorithms to analyze and cluster unlabeled datasets. These algorithms discover hidden patterns or data groupings without the need for human intervention. Common unsupervised learning algorithms include K-Means Clustering and Principal Component Analysis (PCA) ( Hotelling, 1933 ). K-Means Clustering is a method used to categorize unlabeled data into different groups or “clusters” and PCA is a dimensionality reduction method used to reduce the number of input variables in a dataset.

2.3 Reinforcement learning

RL is an area of ML where an agent learns to behave in an environment, by performing certain actions and observing the results or rewards of those actions ( Sutton, 1988 ). The goal is to learn a series of actions that maximize the final reward. Prominent examples of RL algorithms include Q-Learning ( Khenak, 2010 ) and state-action-reward-state-action (SARSA).

Beyond traditional machine learning techniques, the integration of supervised, unsupervised, and reinforcement learning methods with deep learning (DL) and neural networks (NNs) has become significantly popular in the past decade, transforming multiple areas of artificial intelligence (AI). Deep learning, in particular, deserves special attention due to its recent advancements and numerous achievements within the field of computer science. Currently, many researchers are adopting deep neural networks for their specific applications, regardless of whether the problems are supervised, unsupervised, or related to reinforcement learning (RL) ( Hatcher and Yu, 2018 ). Deep learning, a subfield of ML, leverages NNs with three or more layers. These networks attempt to simulate the behavior of the human brain to “learn” from large amounts of data. While traditional ML techniques are often handcrafted, DL models are capable of automatic feature extraction from raw data, making them highly effective and versatile. Convolutional Neural Networks (CNNs) ( Russakovsky et al., 2015 ), a specialized kind of NN, can be trained using supervised learning techniques to identify objects within images recognizing intricate patterns. Similarly, Recurrent Neural Networks (RNNs) and Long Short-Term Memory networks (LSTMs) ( Sutskever et al., 2014 ) excel in sequential data tasks like speech recognition and text translation. The advent of DL has not only provided powerful tools for tasks previously mentioned, such as image classification, text completion, and game playing, but it has also opened doors to more complex problem-solving scenarios that were previously challenging or even impossible to address. The success of DL can be attributed to factors such as the availability of large, labeled datasets, increased computing power, and the development of advanced training techniques, altogether making DL a vital part of the modern ML. In the diverse landscape of ML methodologies, there are numerous algorithms, each with its own strengths and applications. Figure 2 illustrates the list of commonly used algorithms and architectures within the paradigm, capturing techniques from traditional statistical models to the more recent advances in DL.

Figure 2 . Classes of machine learning models.

2.4 Training, validation, and testing

In machine learning, the training, testing, and validation procedures are fundamental steps to develop, evaluate, and refine predictive models. During the training phase, the model learns to make predictions or decisions based on a given dataset, adjusting its parameters to minimize the difference between its predictions and the actual outcomes. The validation phase involves using a separate part of the dataset (the validation set) to fine-tune model parameters and prevent overfitting, which occurs when a model learns the training data too well and fails to generalize to new data. This step is used for selecting the best model version that performs well on unseen data. Usually, the data is separated into two sets: training and testing datasets. It is typical to separate the training data again into several parts (say k parts), and one part is reserved for validation and the rest is used for training. The process is repeated taking each portion of the data as validation set. This is normally referred to as k-fold cross validation ( Wong and Yang, 2017 ). Finally, the testing phase uses the test dataset, which is distinct from the training dataset, to evaluate the model's performance, providing an unbiased assessment of how well the model generalizes to new, unseen data. This structured approach ensures the development of robust, accurate, and generalizable machine learning models.

To identify the performance of a model various metrics and parameters are used in the industry. A confusion matrix is one such performance evaluation tool popularly adopted in machine learning, representing the accuracy of a classification model. It displays the number of true positives, true negatives, false positives, and false negatives. By calculating the TPR and TNR as in Equations 1 , 2 , this matrix aids in analyzing model performance and identifying misclassifications ( Fawcett, 2006 ).

The value TPR and TNR can be calculated using Equations 1 , 2 respectively.

TP (True Positive): The count of instances accurately classified by the model as belonging to the positive class, when they actually are in the positive class.

FP (False Positive): The count of instances incorrectly classified by the model as belonging to the positive class, when they actually are in the negative class.

FN (False Negative): The count of instances incorrectly classified by the model as belonging to the negative class, when they actually are in the positive class.

TN (True Negative): The count of instances accurately classified by the model as belonging to the negative class, when they actually are in the negative class.

3 Key performances requirements of power system protection

Power system protection and controls are critical components of the electrical power grid infrastructure. Power system protection involves deploying a set of strategies and devices designed to detect and isolate faults in power systems, thus minimizing the impact of such faults on the rest of the system. Protection control strategies, on the other hand, are a set of measures initiated to counteract severe disturbances, prevent system collapse, and enable quick recovery to stable operating conditions. Importantly, the term protection does not explicitly indicate that the protective equipment can anticipate or prevent failures; the protective structures designed do not anticipate problems.

Protective relays act only after an event of intolerable conditions and their objective is to minimize the duration of the problem, limit damage, reduce downtime and other problems created by the event. In asset protection, this task is performed by circuit breakers controlled by protection relays. They isolate areas or problematic elements on the circuit. These actions can be divided into two groups: primary and back-up. Primary protection isolates the faulty equipment with exceptional speed and precision, while backup protection acts as fail-safe, clearing faults missed by the primary system. Backup protection is slower, but covers a wider area, and its settings must be carefully calibrated to adapt to varying system conditions ( Phadke et al., 2016 ). Based on these principles, it is clear that protection systems must be fast enough and selective enough to isolate faults without affecting the entire network, thereby improving power system reliability. On the other hand, system protection response to abnormal operating conditions affecting a wider area. In both cases, the protection schemes should avoid excessive complexity. Overall, the system should prioritize simplicity and effectiveness while remaining economically viable ( IEEE, 1988 ; CIGRE, 2001 ). Although traditional protection methods have been well established over the last century, the use of machine learning algorithms as support can enhance key performance requirements of such protection schemes.

The traditional protection and control strategies must ensure five principles: reliability, selectivity, speed, simplicity, and economy, to be considered as effective and efficient ( Blackburn and Domin, 2006 ), and machine learning algorithms must contribute to the fulfillment of these key performance requirements.

3.1 Reliability

Reliability is defined on top of two concepts: dependability and security. Dependability is defined as the degree of certainty that the relay will operate correctly. Security is the degree of certainty that the relay will not operate incorrectly.

3.2 Selectivity

Protection relays have a designated protection, while also offering delayed backup protection for adjacent zones. The selectivity is a key requirement to minimize the extent of outages during fault events and it is an area where traditional protection struggles due to use of simple decision functions with a limited set of inputs. Machine learning that can be effectively used to improve the selectivity of difficult protection problems by recognizing patterns and complex scenarios.

In power system protection, rapid fault isolation is desirable, but achieving very high-speed operation can lead to undesired actions. Time remains a reliable means of distinguishing tolerable from intolerable transients. It is desirable for a protection relay to operate as soon as a fault is correctly detected. However, due to the operating principles of traditional protections, there is a trade-off between speed and false positive detections. Machine learning algorithms can mitigate this by leveraging different detection principles and input features that can help do an early detection of the fault ( Azhar et al., 2022 ).

3.4 Simplicity

The design of a protective relay system should give priority to simplicity and straightforwardness, while still achieving the intended objectives. Each additional unit or component that enhances protection but is not essential to the basic protection requirements should be carefully considered. Each added element introduces a potential source of problems and increased maintenance. Incorrect operation or unavailability of protection can lead to catastrophic problems in an electrical system, since problems in the protection system can significantly affect the entire system, possibly more than any other component. Simplicity gains relevance when considering machine learning algorithms for power system protection. Machine learning algorithms often exhibit nonlinear decision boundaries that can cause incorrect classifications, even when the overall performance of the algorithm is satisfactory ( Huang W. R. et al., 2020 ).

One of the most common shortcomings of machine learning based protections when compared to traditional protections comes from model interpretability. Traditional protections follow clear physical principles that are well-understood. This makes it possible to stack multiple components together in a meaningful way. Machine learning algorithm interpretability is a whole area that deals with this issue and aims to explain what is often thought of as black-box algorithms that are inherently complex ( Molnar, 2020 ). On the other hand, machine learning algorithms can be constrained to operate within a bounded region of well-understood physical quantities e.g., a region of the R-X plane. Constraining an inherently complex machine learning algorithm allows it to be stacked on top of a simpler but protection principle without increasing the whole system complexity.

3.5 Economics

The balance between maximum protection and cost-effectiveness is critical. Initial savings may tempt one to choose the least expensive protection system. However, this can lead to reliability issues, installation challenges, and higher maintenance costs. The cost of protection may seem high up front, but it pales in comparison to the potential cost of equipment damage and downtime resulting from inadequate protection. Prioritizing proper protection at the outset is wiser than cutting corners and paying more later.

Reliability, selectivity, speed, simplicity, and economics are crucial for providing uninterrupted power supply and reducing the risk of power failures and outages. In the current context of machine learning algorithms insertion, having robust scheme must enhance these five key principles.

4 Opportunities for machine learning in power system protection

Opportunities for ML in Power System Protection are vast and continue to grow with the technological advancements in both fields. Fundamentally, the machine learning model is capable of making decisions that can replace the logics in conventional protection systems, or assist the logical functions to make better decisions, as illustrated in Figure 3 as applicable to asset protection. In some cases, it is possible for conventional decision-making processes to operate in parallel with machine learning models.

Figure 3 . Comparison of Conventional and Machine Learning-Based Relays. Panel (A) shows the operation of a conventional relay based on logic, while panel (B) illustrates a relay that uses machine learning for decision-making.

There are many power system protection and control functions that can be improved by using ML techniques ( Rajapakse et al., 2002 ; Zhou et al., 2010 ; Jayamaha et al., 2019 ). These areas offer a rich landscape for innovation, where ML can contribute to developing new solutions and improving existing methodologies. In this section, potential application areas are explored, ranging from power system stability to emergency control, mis-operation detection, and more.

There is a vast amount of literature covering potential application areas of ML techniques in power system protection and emergency control. Therefore, a selection criterion was designed as illustrated in Figure 4 to select several representative research papers for each application. Initially, this selection criterion selects all the research papers that contain the following metadata: (1) machine learning techniques such as traditional machine learning (i.e; SVM, DT...etc.), deep learning and reinforcement learning, (2) Potential power system protection and emergency control applications such as power system stability, high impedance fault detection…etc. Then a pool of research papers was created by using IEEE Xplore ( https://ieeexplore.ieee.org ), Scopus ( https://www.scopus.com ) and Google Scholar databases. This pool comprised of candidate research papers which were peer-reviewed (journal and conference papers), contained specific key words and published within a time range of 2004–2024.

Figure 4 . Research paper selection criterion.

Figure 5 shows the percentages of different papers categorized based on the application. Still there were considerable number of papers under each application. Therefore, a limited number of papers were selected manually considering the quality, number of citations and diversity of algorithms utilized, to include in this review.

Figure 5 . Percentages of pooled research papers categorized based on the applications.

4.1 Power system stability

Maintaining the stability of the power system is one of the main objectives of a power system protection and control system. Generally, persisted contingencies or multiple contingencies occur in power systems beyond the designed tolerance level of protection and control systems is the root cause of power system instabilities and blackouts. Power system stability can be mainly categorized into voltage stability, rotor angle stability, frequency stability, oscillatory stability, and inverter driven stability ( Hatziargyriou et al., 2020 ). Most of these stability phenomena occur within several seconds which is not feasible to detect using model driven approaches in real-time. Therefore, many researchers have drawn their attention on ML based data driven approaches which can predict the power system stability status within few milli-seconds with an acceptable accuracy. In this section, several ML-based assessment and control action design approaches are discussed under each stability category.

4.1.1 Voltage stability

The voltage stability of a power system can be defined as the ability of the system to maintain a steady voltage close to nominal value at all buses after the system has been subjected to a disturbance. Conventional voltage stability assessment methods such as continuation power flow (long-term voltage stability) and transient stability analysis (short-term voltage stability) lack the applicability for large power systems in real-time due to high computation time and model dependency. Recently, there are many voltages stability assessing approaches proposed in literature based on ML due to their data driven nature. These approaches use both shallow and DL algorithms. Assessing short-term voltage stability of a power system is a time sensitive process which requires trajectory feature identification within few seconds therefore feature engineering approaches can be seen in Zhu et al. (2016 , 2020) , Yang et al. (2018) , Dharmapala et al. (2020) , and Dharmapala and Rajapakse (2024) . These features are Time Series Shapelets (TSS) in Zhu et al. (2016) and Zhu et al. (2020) , pre-identified templates in Dharmapala and Rajapakse (2024) and online induction motor slip in Yang et al. (2018) . On the other hand, deep learning algorithms such as LSTM can grasp features without pre-processing (feature computation). LSTM based SVS assessment schemes can be found in Zhang et al. (2021) and Zhu et al. (2021) . A loadability prediction scheme to assess long-term voltage stability is proposed and validated using real-time data from phasor measurement units (PMUs) in Dharmapala et al. (2020) . A summary of utilized ML algorithms in these approaches is tabulated in Table 1 .

Table 1 . Summary of ML based voltage stability applications.

4.1.2 Rotor angle stability

The ability of the interconnected synchronous machines in a power system to remain in synchronism under normal operating conditions and to regain synchronism after being subjected to a small or large disturbance is defined as the rotor angle stability ( Hatziargyriou et al., 2020 ). Generally, rotor angle instabilities occur within few seconds. Therefore, rotor angle stability is another area where data driven ML based approaches are frequently applied. In Amjady and Majedi (2007) , the power system is partitioned into subsets each including two or more synchronous machines using the data from each partition. ANN is trained and final decision is obtained through a voting mechanism. Proximity to pre-identified voltage templates are used as features to a SVM classifier in Rajapakse et al. (2010) to differentiate rotor angle stability/instability. In Gomez et al. (2011) , direct voltages measurements are input to a SVM classifier to predict the transient stability status after a disturbance. A small signal stability assessment scheme is proposed in Dorado-Rojas et al. (2021) by utilizing deep learning CNN. Time and frequency domain measurements have been utilized to extract features in Kamwa et al. (2009) and fuzzy rule base is used to improve the decision boundary tuning. Similar to deep learning-based voltage stability assessment applications, these deep learning approaches use sequential measurements without feature identification. Some of these ML approaches found in literature are tabulated in Table 2 .

Table 2 . Summary of ML based rotor angle stability applications.

4.1.3 Frequency stability

Frequency stability refers to the ability of a power system to maintain steady frequency level following a significant imbalance between severe system upset resulting a significant imbalance between generation and load ( Hatziargyriou et al., 2020 ). Frequency instabilities initiate in the form of sustained frequency swings or large frequency deviations and can be led to tripping of generators and/or loads. Due to the fast response, it is difficult to differentiate stable frequency swing from unstable frequency swing; however enhancement of ML techniques enables the power system frequency stability assessment and control applications. In Behdadnia et al. (2021) , a practical PMU measurement-based frequency stability analysis scheme which can detect and eliminate erroneous measurement is proposed. The main factors that contribute to the power system frequency response is considered when training the ML model in Bo et al. (2014) and Xie and Sun (2021) . Bandwidth requirement of PMU data for frequency stability analysis is reduced in method proposed in Tripathi (2018) . Few of the frequency stability applications found in literature are tabulated in Table 3 .

Table 3 . Summary of ML based frequency stability applications.

4.1.4 Oscillatory stability

The resonance, in general, occurs when energy exchange takes place periodically in an oscillatory manner. These oscillations grow in case of insufficient dissipation of energy in the flow path and are manifested (in electrical power systems) in magnification of voltage/current/torque magnitudes ( Hatziargyriou et al., 2020 ). Oscillatory stability monitoring and controlling is a highly researched area, but application of AI is relatively limited. Eigenvalue region prediction of critical stability modes results from inter-area oscillations is proposed in Teeuwsen et al. (2006) . A scheme which provides offline training, update and online predicting is proposed in Liu et al. (2021) . In Cepeda et al. (2022) , deep learning-based assessment scheme is proposed which only uses system frequency as the input. An optimal control strategy using thyristor-controlled series capacitors (TCSC) to damp power system oscillations is proposed in Ernst et al. (2004) and in this method Reinforcement Learning (RL) is utilized to obtain the control policy. A summary of various ML approaches found in literature for oscillatory stability monitoring and controlling is tabulated in Table 4 .

Table 4 . Summary of ML based oscillatory stability applications.

4.1.5 Inverter driven stability

The dynamic behavior of Inverter Based Generators (IBGs) is clearly different from conventional synchronous generators. Typical IBG relies on control loops and algorithms with fast response times, such as the PLL and the inner-current control loops ( Hatziargyriou et al., 2020 ). In this regard, the wide timescale related to the controls of CIGs can result in cross couplings with both the electromechanical dynamics of machines and the electromagnetic transients of the network, which may lead to unstable power system oscillations over a wide frequency range. The research found in literature on inverter driven stability is mainly focused on controller optimization using ML techniques such as RL. In Gheisarnejad and Khooban (2020) , a control scheme based on Active Disturbance Rejection Controller (ADRC) is implemented with optimal setting training process. A data-driven optimal control strategy for Virtual Synchronous Generator (VSG) operation is proposed in Li Y. et al. (2021) with control targets of maintaining frequency within operating limits, damping oscillations, and smoothing frequency response. A summary of various RL approaches to inverter controller tuning that found in literature is tabulated in Table 5 .

Table 5 . Summary of ML based inverter driven stability applications.

4.2 Communication infrastructure for protection and control

IEC 61850 standard ( Report, 2014 ) is developed in 2014 to standardize power system communication network and automate systems. Because of the standard object-oriented approach, it allows interoperability between devices regardless of different vendors. IEC 61850 is the protocol used by power system protection and control devices for communication. Generic Object-Oriented Substation Event (GOOSE) and Sample Values (SV) are two main data protocols proposed in the standard. GOOSE messages are triggered by certain events in the power system and are transmitted to take necessary reactionary precautions, i.e., a trip command is sent to a circuit breaker via GOOSE message after a relay picks up excessive current readings. On the other hand, SV messages carry periodic samples of critical grid parameters such as bus frequency, voltage etc. Due to the critical nature of the places of their use, i.e., measurements for power system protection, frequency, and voltage control. Due to the importance of these data, these communications infrastructures are becoming more vulnerable to cyber-attacks. There are several research that can be found in literature to identify such intrusions using ML. In Ustun et al. (2021a) , based on the frequency and nature of GOOSE messages a ML based approach is utilized to differentiate usual operation from cyber-attacks. An intrusion detection and communication traffic monitoring system based on SV data is proposed in Ustun et al. (2021b) . A SVM based approach to identify compromised devices in a smart grid is proposed in Kaygusuz et al. (2018) . These approaches are summarized and tabulated in Table 6 .

Table 6 . ML based power system communication applications.

4.3 Emergency control

Modern power systems are characterized by a significant increase in uncertainties and the risk of major outages due to their operation close to limits, decreased inertia, and the integration of renewable energy sources with fluctuating output. This necessitates a rethinking and enhancement of emergency control strategies, which have traditionally been designed offline based on worst-case or typical operational scenarios. One of the promising approaches to tackle these challenges is to implement a RL based agent which can interact with the power system environment ( Ernst et al., 2004 ). These tools can harness real-time data from multiple sources to provide accurate, near-instantaneous assessments of the grid's status. Such models could not only forecast potential threats but also suggest optimal reactive measures in real time, enhancing the adaptability and robustness of the system. Moreover, the development of distributed energy resources such as solar plants, wind plants, and battery storage systems, provides an opportunity to create localized emergency control strategies. By allowing areas of the grid to function independently during crises, these resources can help mitigate the risk of widespread blackouts. In this section, we explore a variety of ML methodologies that are employed to efficiently handle the complexities of emergency control problems.

An innovative approach introduced in Li et al. (2022) utilizes an autonomous control method based on the DDPG to effectively mitigate issues associated with under-voltage load shedding. Also adaptive under voltage load shedding and emergency control schemes using deep reinforcement learning (DRL) is proposed in Huang Q. et al. (2020) . Similarly, Chen et al. (2021) presents a model-free emergency frequency control strategy, leveraging reinforcement learning to navigate the complexities of such scenarios. Notably, it introduces a multi-Q-learning-based method to limit the number of emergency scenarios that should be addressed when designing the system. To learn the optimal solutions for identified general emergency scenarios, DDPG algorithm is adopted. In a similar vein ( Vu et al., 2021 ), proposes an emergency load shedding technique aimed at enhancing the voltage recovery of post-fault conditions. This technique notably incorporates safe RL to ensure the reliability and safety of the implemented solutions. Collectively, these studies emphasize the growing support on advanced RL techniques in developing responsive control mechanisms for emergencies and illustrating a shift toward self-adjusting systems in power grid management. Table 7 gives a summary of ML based methods to solve the emergency control problems.

Table 7 . Summary of ML based methods to solve the emergency control problem.

4.4 Mis-operation of protection systems

Mis-operation in a part of the system happens when it doesn't perform as planned or functions outside its assigned area of protection. When a protection system either fails or operates improperly, it leads to a less stable state. This not only disrupts the safeguarding of the system's equipment but also contributes to outages in transmission and negatively impacts the overall reliability of the system. The Subcommittee for Protective Relay at MRO has analyzed mis-operation submissions from the period between 2010 and the initial quarter of 2016, categorizing them as shown in Figure 6 ( MRO Protective Relay Subcommittee, 2017 ). The leading three causes identified are connected to mistakes in logic, faults in design, and errors made by personnel that were left uncorrected. Additionally, there were Mis-operations traced back to specific failures in the relay system and errors in the DC system.

Figure 6 . Breaker failure mis-operation per cause code.

Mis-operation in power system protection is a grave concern, leading to the development of numerous techniques and practices to reduce its occurrence. In the WECC report (Western Electricity Coordinating Council, 2018) , Mis-operations Reduction Strategies are classified into seven distinct subjects, and some of the report's recommendations are provided below.

• Ground overcurrent protection: Coordination with system changes and contingencies is vital when designing ground instantaneous overcurrent (50G). The 50G should be set higher than the maximum external fault current plus an additional margin. Moreover, due to the variability in fault levels, coordination studies are necessary, and the impact of mutual coupling must be taken into account during the design of Ground time-overcurrent 51(G).

• Human performance during commissioning : A comprehensive commissioning process is recommended, with specific practices outlined to identify errors before the energizing of new equipment.

• Knowledge transfer : It is essential to have a well-documented plan for sharing knowledge.

• Limited Information for Investigations

• Root cause analysis : Employing staff members trained in root cause analysis to conduct event investigations can enhance the examination beyond the obvious cause, identifying hidden errors elsewhere in the system.

• Settings validation: Despite the development of many techniques to reduce mis-operations, no sufficient method has been established to detect mis-operation in real time. Consequently, analyzing and finding solutions from ML techniques remains an open field for researchers. However, some work related to mis-operation detection based on PMU data analyzing is proposed. A tool has been implemented to detect real-time mis-operations in transmission line relays as described in Esmaeilian et al. (2015) . This tool utilizes time-synchronized measurements gathered from both ends of the line during disturbances. The proposed methodology effectively confirms whether the line tripping was due to a mis-operation of protective relays. The paper referenced in Banerjee et al. (2019) explores methods for enhancing real-time situational awareness of power systems. It utilizes PMU data to identify dynamic events occurring within the system. Additionally, the paper proposes various options for a data-driven model and investigates the performance of certain patterns in classifying PMU disturbance data. A summary of approaches used for mis-operation detection is given in Table 8 .

Table 8 . PMU data-based methods to identify the relay mis-operations.

4.5 High impedance fault detection

Unlike most power system line faults, High Impedance Faults (HIF) prevent the generation of sufficient current that is required to trip the over-current relays due to high grounding impedance. These faults could jeopardize human safety by unintentional contact with an energized exposed conductor. Hence it is an important but difficult task to detect these types of faults. Therefore, there are specially designed algorithms for HIF detection. In HIF detection, after feature extraction from the measurements a boundary should be found to separate a faulty state from healthy ones. Among different classification methods ML-based methods have higher accuracy in pattern classification, fast response, noise removal ability and prediction capability ( Hotelling, 1933 ). An approach which uses TT- transform to extract time-time distribution features is proposed in Nikoofekr et al. (2013) . In Moravej et al. (2015) , Dual Tree- Complex Wavelet Transform (DT-CWT) is utilized for extraction features and then fed to the NN classifier. Deep neural networks have been utilized in Rai et al. (2021) and Sirojan et al. (2022) to grasp unique characteristics at an event of HIF and different such event from power system switching operations and measurement noises. A summary of these ML approaches is tabulated in Table 9 .

Table 9 . Summary of ML based high impedance fault detection.

4.6 Photovoltaic systems and inverter based generation

Over the last two decades, the growing number of inverter-based generation (IBG) resources has brought multiple challenges for power system protection. The variable nature of wind and solar power, the electrical characteristics of inverter circuits and the growing number of generators in distribution systems make protection challenging from the optics of conventional methodologies. Emerging challenges have been widely identified and studied; nevertheless, reliable fault detection, classification, and localization are still a matter of discussion. The complex nature of phenomena that arise from the growing penetration of renewable and IGB resources has prompted researchers to employ ML techniques to address these issues ( Wischkaemper and Brahma, 2021 ). As stated before, protection issues have been widely identified, and photovoltaic (PV) systems are no exception. In Alam et al. (2015) , the authors thoroughly review the types of faults affecting PV arrays, the methods used to detect them, and their shortcomings. The authors identify three types of faults in PV arrays: line-to-ground, line-to-line, and arc. Moreover, they identify challenging issues for conventional protections: the line-to-ground blind spot, line-to-line faults under low light conditions, high impedance paths inside the array, and the detection of arc faults. On the other hand, conditions such as open-circuit and partial shading are identified as disturbances that need to be reliably detected. The wide variety of array shapes and sizes make it necessary to design algorithms with normalized input features that can work reliably at any scale; normalization of electrical quantities is usually done with respect to V OC and I SC ( Yi and Etemadi, 2017 ) and ( Kumar et al., 2023 ) uses the rate of change of the conductance and the fill factor. Pure measurements are used in Zhao et al. (2015) , Yi and Etemadi (2017) , Madeti and Singh (2018) , and Kumar et al. (2023) use feature engineering for trans-forming the original measurements into quantities that can improve the algorithm. In contrast, series arc faults in PV systems produce very little changes in voltage and current; detection of series arc faults in PV systems can be made using advanced signal processing techniques that can detect signature waveforms produced during the fault; this approach is adopted in Kumar et al. (2023) using a DL algorithm. A summary of ML based PV system protection approaches is provided in Table 10 .

Table 10 . ML based protection for PV arrays.

The proliferation of IBG has changed the topology of power systems by allowing generation resources to be located on the demand side and connected to distribution circuits as microgrids or active distribution systems. Protection issues and challenges have been studied and identified ( Kumar et al., 2023 ): low fault currents, protection blinding, sympathetic tripping, and islanding detection. Moreover, unlike passive distribution grids, the distributed nature of this type of generation makes power system protection a shared responsibility, as upstream protections may not be enough to guarantee a safe operation, e.g. the need to cease to energize during islanding. Unintentional islanding is a situation that arises when an active distribution system is disconnected from the bulk power system. Fast and reliable islanding detection has proven to be challenging for traditional protection methodologies as these are heavily dependent on the operating state before the fault or require additional infrastructure that makes them expensive and complex ( Lidula et al., 2009 ). Islanding detection can be carried out by leveraging signal processing techniques that extract features from voltage and current signals, these features can be used to train classifiers that can improve detection. Particularly, features derived from the discrete wavelet transform (DWT) decomposition of voltage and current signals carry the necessary information to successfully detect islanding in active distribution grids ( Lidula and Rajapakse, 2010 , 2012 ). Several ML based approaches proposed for microgrid related protection functions are presented in Table 11 .

Table 11 . ML based protection for microgrids.

4.7 Transformer inter-turn faults

Faults in power transformers can lead to interruptions, equipment damage, or even issues with the stability of the entire system. Short circuits in a few turns, known as inter-turn faults, generate a fault current among the involved windings that can cause thermal overload in the region and create other conditions such as phase-phase faults, phase-ground faults, and over-fluxing ( Subramanian, 2020 ). Thus, it is important to have a comprehensive protection scheme for power transformers. Although tradition-al challenges such as inrush current, core saturation, external faults, etc. affect it, conventional protection relay, based on elements such as current differential, Buchholz, Volt/Hz, and overcurrent, has been widely used for many years ( Pani et al., 2020 ). The frequency response analysis (FRA) also is a method widely used to recognize the changes on the winding impedance changes after an internal fault occurs ( Khalili Senobari et al., 2018 ). The main drawback of this technique is lack of consistency in analysis, as there is no universally accepted code for interpreting FRA ( Li Z. et al., 2021 ). The use of digital image processing methods for FRA interpretation has been the focus in Aljohani and Abu-Siada (2016) and Vosoughi and Samimi (2022) . ML techniques, such as decision trees ( Bigdeli et al., 2021 ), SVM ( Liu et al., 2019 ), and ANN ( Behkam et al., 2022a , b , c ), have effectively been used to interpret the FRA signatures for classifying types of faults. This enhances the precision of fault diagnosis by minimizing errors in subjective analysis. Voltage and current waveforms from the transformer terminals or magnetization inrush current are some of the features used by researchers to classify and identify internal faults. DL frameworks, including classification autoencoders, have been employed by some researchers ( Duan et al., 2019 ). Additionally, other ML techniques such as decision trees, random forest, and gradient boost classifiers ( Simões et al., 2021 ), have been applied to analyze differential current.

4.8 Post-event analysis

Due to the increased demand and the integration of microgrids and renewable energy, the complexity of power systems has reached a level that makes it challenging to detect events, estimate stability issues, and prevent potential blackouts. This behavior is caused by non-linear higher-order elements and the time-varying nature of power grids, demanding efficient algorithms for dynamic monitoring, control, and protection ( Zhang et al., 2011 ; Thomas et al., 2020 ). PMUs have been essential in studying post-fault anomalies during significant events worldwide. However, their extensive distribution has raised challenges upon scaling up data analysis ( Dahal et al., 2014 ). Despite the abundance of high-resolution PMU data, the optimization of PMU data to predict anomalies is still a developing topic, which highlights the need to apply supervised learning techniques. Long periods of data can be analyzed in order to detect patterns and potential anomalies in the system and characterize its behavior in response to external factors ( Hou et al., 2020 ). One of the analyzed strategies involves event detection and clustering to form a hierarchy of events for PMU data. Classification tools, such as continuity, correlation, SNR, among others, give analysts a means for event detection and inputs for case simulations ( Hou et al., 2020 ). Some of ML based approaches proposed for post-event analysis are presented in Table 12 .

Table 12 . ML based post-event analysis.

5 Exploring the bottlenecks in applying ML to power system protection

Despite the substantial progress of applying ML to power system applications over the past decade, it's surprising to observe that ML isn't popular in practical applications in power system protection. In Wischkaemper and Brahma (2021) , it is mentioned that there isn't a single commercial relay that uses ML for either primary or backup protection at the time. This chapter delves into the complex challenges involved in integrating ML into power system protection applications.

5.1 Lack of interpretability

A central challenge lies in the lack of interpretability of modern ML models. These models, particularly complex ones like deep neural networks, are frequently considered “black boxes,” meaning it's hard to grasp why they make certain decisions ( Zaker et al., 2013 ; Rojas-Dueñas et al., 2020 ; Aghababaeyan et al., 2023 ). This lack of clarity becomes critical when dealing with power systems, where every decision carries far-reaching consequences. A single misstep could potentially trigger a system-wide blackout, causing significant damage to a region's economy ( Yamashita et al., 2009 ; Alhelou et al., 2019 ). Hence, the need for transparency and explainability in decision-making is vital, an aspect that current ML models often fail to meet.

5.2 Lack of guaranteed performance

Another significant concern in the application of ML to power system protection is the lack of guaranteed performance. Traditionally, the performance of ML models is established through “testing and validation.” But one must question if this standard method of testing is adequate for crucial applications such as power system protection and control. Consider an ML model that demonstrates an accuracy of 99% during testing. While this might seem commendable in many fields ( Wu and Chen, 2016 ; Islam et al., 2018 ) in power system applications, a 1% error margin could result in severe consequences. This seemingly small percentage of error could provoke a massive system failure, emphasizing the severity of even minor inaccuracies. Therefore, ensuring safe operation in the relevant region is essential. This safety factor is inherently incorporated into traditional methods. These established methods function safely and optimally within certain operating conditions, and while they might operate less efficiently outside these limits, they maintain a safe mode of operation. ML models, if they are to be utilized in power system protection, must follow these same rigorous standards of safety.

Moreover, a fundamental issue lies in the methodology used to test ML models. Typically, these models are validated using a subset of data from the entire dataset. Although a model may exhibit a perfect accuracy of 100% for this specific subset, it doesn't assure a similar level of performance for the entire continuous region of operation. Practically speaking, testing an ML model across the entirety of this continuous region is not feasible, given the infinite number of data points it encompasses. This adds another layer of complexity to the problem, presenting the need to develop a robust method that can validate the model's performance across the entire operating region, rather than just a data subset.

5.3 Scarcity of high-quality data

The scarcity of high-quality data in the power system sector presents a formidable challenge when developing data driven technics. ML models succeed on high-quality data, with their performance directly dependent on it ( Sessions and Valtorta, 2006 ). However, unlike the field of computer engineering, where ML models are traditionally applied, the power system industry relies on tangible sensors such as current and voltage transformers. The data these sensors generate can contain noise and may be imbalanced, both factors that could deter model performance ( Vega et al., 2007 ). Additionally, the necessary infrastructure for effective data collection often falls short. Furthermore, although data is collected and stored, much of it remains inaccessible to the wider research community due to security and privacy constraints.

5.4 Uneven datasets

Adding to these difficulties is the issue of an uneven balance between normal and abnormal data. Power systems typically operate in a 'normal' state, meaning that 99.9% of stored data reflects this healthy state. Yet, to construct a robust ML model, exposure to data representing abnormal or alert states is crucial. This lack of 'abnormal' data scenarios forces the model to 'unlearn' most of the time, with only a few in-stances providing opportunities for new learning. Also, the uneven dataset for training led to biases in the ML model ( Mehta et al., 2019 ). Consequently, developing a ML model solely from practical sensor data is challenging. To overcome these hurdles, combining practical sensor data with simulated data may be necessary for developing a robust model. This approach could help to counterbalance the issues arising from data quality, accessibility, and imbalance, ultimately improving the model's ability to predict and react to both normal and abnormal states within power systems.

5.5 Adversarial challenges

An additional obstacle lies in dealing with adversarial examples. These are specifically crafted inputs that aim to deceive a NN, leading to incorrect classification of a given input ( Goodfellow et al., 2015 ). The presence of adversarial examples poses a significant challenge for the successful application of ML to power system protection. In the context of power systems, an adversarial example could be a seemingly changed input designed to induce misclassification ( Chen et al., 2018 ). For instance, an adversarial example might make a normally operating power system appear to be in a state of fault, or it could mask an actual fault, making it seem like the system is operating under normal conditions. The successful identification and handling of such instances is vital to avoid potentially disastrous consequences, such as shutdowns or system failures.

5.6 The curse of dimensionality

The challenge of scalability is another significant factor to consider when applying ML to power system protection. In the context of analyzing a complex power grid, the volume of data to be processed can be substantial. As a result, the state and decision spaces, which are the sets of possible states and actions respectively, can increase dramatically with the number of elements in the grid. This issue is referred to as the ‘curse of dimensionality'. The curse of dimensionality presents a unique set of challenges. It becomes increasingly difficult to efficiently analyze and process data as the size of the power grid grows. A ML model might perform excellently on a smaller scale but struggle as the number of features increases. The exponential growth of possible states and actions can quickly overwhelm computational resources and lead to extended processing times.

5.7 Integrating ML models into existing power systems

The integration of ML models into existing power system infrastructures brings its own set of challenges. These power systems, often built and enhanced over many decades, rely on deterministic engineering principles. As such, they incorporate detailed requirements and safeguards to ensure robust and reliable operation. ML, by contrast, is fundamentally a probabilistic approach, and its integration into these deterministic systems can lead to significant technical and operational complications. On a technical level, many existing systems were not designed with ML integration in mind. Implementing such models might require extensive modification or even a complete redesign of current systems. Moreover, ML models typically demand substantial computational resources that the existing hardware might not be able to provide. An additional challenge arises from the lack of established regulations. As the application of ML in power system protection is relatively novel, comprehensive standards and guidelines have not yet been defined. This situation can cause uncertainty about compliance and safety, complicating the integration of ML into power systems even further.

6 Future direction

The prediction of the future of ML technology is challenging, especially considering its rapid growth. It's evident that AI and ML will play a significant role in the development of future applications. While these technologies are not yet mature enough to be directly applied to power system protection, there are certain promising innovations within the field that are demonstrating substantial potential and undergoing consistent development. There should be significant attention required to overcome the barriers mentioned in Section 5 to bring the ML models to real practical applications. In the following sections, there is an examination of these new advancements, along with an exploration of how they might be applied in power system protection. Figure 7 presents a summary of possible future advancements and their potential to address the bottlenecks/challenges presented in Section 5.

Figure 7 . Challenges and solutions in applying machine learning models to power system protection.

6.1 Power system protection with explainable ML models

Explainability or interpretability is a prerequisite to ensure the scientific value and safety of the outcome ( Doshi-Velez and Kim, 2017 ). In this context, research directions such as explainable ML models have emerged over the past decade ( Ribeiro et al., 2016 ; Došilović et al., 2018 ). The idea is to make the decision-making process of the ML models understandable by humans, providing clear explanations for their predictions. This will be a solution to when it comes to current barriers of applying deep neural network to practical applications. If deep neural network's black box nature is clearly interpretable, there will be trust among the communities to apply it to critical applications such as a power system ( Machlev et al., 2022 ). With the advent of these explainable ML models, the future of power system protection looks promising. The use of these models can bring about enhanced system protection, optimized operational efficiency, and increased reliability. There are some works that have started to emerge in the power system such as emergency control ( Zhang et al., 2022 ).

6.2 Use of emerging technologies for protection applications

6.2.1 liquid neural networks.

A groundbreaking type of network known as the liquid neural network, which possesses the ability of continuous learning during its operation, not limited to just the training process, was developed recently ( Hasani et al., 2021 ). It is demonstrated that liquid neural network has the ability to steer an autonomous car with the by processing the data with only 19 neurons and 253 synapses in the network ( Lechner et al., 2020 ). Traditionally this kind of task is achieved from a CNN, and it requires many neurons and synapses to achieve such a task. This method helps to reduce the size of the network which internally helps to understand them and explain the behavior of the NN. The development of this innovative method holds great promise for applications involving decision-making based on dynamic data streams that evolve over time. As a result, this novel approach could be effectively utilized in creating adaptive, compact models for power system protection applications, particularly in scenarios where the topology undergoes continuous changes over time.

6.2.2 Physics-informed neural networks

A physics-informed neural network (PINN) is a type of NN which is trained to solve supervised learning problems while considering any given laws of physics described by non-linear partial differential equations. It is shown that the method is effective for solving some classical non-linear problems in fluid dynamics, quantum mechanics and reaction-diffusion systems ( Raissi et al., 2019 ). PINNs work by enforcing the known physical laws as constraints during the learning process, thereby guiding the network to learn a solution that not only fits the data but also aligns with the underlying physics of the problem. Generally traditional ML models often do not incorporate any prior knowledge about the physical system they are modeling. PINNs, on the other hand, are designed to incorporate physical laws and equations that govern the system as a part of the NN architecture. This is one of the key advantages when applying PINN based ML models to complex, physics-based applications such as a power system.

There are several reasons why PINNs could be particularly useful for power system protection applications. The most important advantage is that improved generalizability. As the PINNs incorporate the fundamental physics that govern power systems, they offer a higher level of generalizability compared to traditional ML models. This offers a reliable protection decision even in scenarios that have not been explicitly encountered during the training phase. There is a significant attention among the power system community to apply PINNs to solve traditional power system problems ( Misyris et al., 2020 ; Bragone et al., 2022 ; Huang and Wang, 2022 ). Nevertheless, the method is not readily applied for power system protection applications. However, as PINNs' predictions are grounded in the known physical laws, which can make their outputs more understandable and reliable to power system engineers. This transparency can lead to higher trust and adoption of these models in future.

6.3 Applying transfer learning to enable knowledge sharing across different ML models

Transfer learning is a concept in ML that includes storing knowledge gained while solving one problem and applying it to a different but related problem ( Ribani and Marengoni, 2019 ). It takes the advantage on the knowledge which obtained from one task to another related task. In essence, transfer learning allows us to leverage pre-existing models that have been trained on large datasets, potentially saving significant computational resources and time in model development ( Zhuang et al., 2020 ). On the other hand, power systems are reliable by design, meaning that events like faults are relatively rare, making it challenging to gather a sufficient quantity of data to train ML models. As the transfer learning allows to leverage the pre-existing models, it could be used to solve the problem of data scarcity ( Pan and Yang, 2010 ).

6.4 Multi-agent reinforcement learning in emergency control

In power systems, emergencies often arise due to unpredictable disturbances like equipment failure, power demand surges, or natural disasters, potentially leading to blackouts. These emergencies require fast and efficient decision-making for control actions to prevent system collapse. Multi agent reinforcement learning (MARL), with its capability to process large-scale multidimensional data and make timely decisions, offers a promising approach to managing these emergencies ( Busoniu et al., 2008 ; Chu et al., 2020 ). MARL involves deploying multiple RL agents across the power system ( Biagioni et al., 2022 ). Each agent focuses on a specific area or component of the system, reducing the complexity of the problem space it needs to manage. In addition, the agent, embedded with Deep RL, takes actions like adjusting generator output, controlling switchgear, or managing demand response to maintain system stability ( Wang et al., 2021 ). This approach can enhance the learning efficiency and decision-making ability of the overall system. However, the coordination between multiple RL agents is complex, as changes in one area may affect the others ( Canese et al., 2021 ). Advanced methods and communication protocols need to be developed to ensure the efficient operation of the overall system. Looking forward, as the power systems continue to grow and become more integrated with renewable energy sources, the complexity of managing these systems under emergency conditions will only increase. However, with the continuous advancements in RL and its ability to learn and adapt from experience, MARL provides a robust toolset for future power system protection and emergency control.

6.5 Power system protection with generative AI

As we look toward the future of power system protection, the integration of generative AI will play a significant role in the future. This advanced form of artificial intelligence, capable of generating new data and models, promises to revolutionize the way power systems are monitored and protected. These AI models could simulate a vast range of potential scenarios, including rare and complex fault conditions, enabling the development of more robust protection strategies. Additionally, the adaptive nature of generative AI means that protection systems could continuously evolve in response to changing grid conditions and emerging threats, such as cyber-attacks or extreme weather events.

7 Conclusions

In conclusion, this paper has underlined the importance of traditional methods in power system protection, methods that have been transparently and robustly developed over the past century. Given the high cost associated with protection failures, these traditional methods do not need to be completely replaced if they continue to follow the five principles outlined in Section 3. Nevertheless, clearly there exists room for enhancement within these conventional methods, and in specific scenarios, ML techniques may offer valuable augmentation to classical approaches. The paper has emphasized potential applications of ML that tackle challenges in power system protection that are difficult to overcome with traditional methods, as detailed in Section 4. Despite the considerable number of research and development efforts that have been reported to date, many of these efforts are still a long way from being translated into practical applications. This paper concludes that the use of ML technology for power system protection is still in its early stages. It identifies significant barriers to implementing ML in power system protection, as described in Section 5. Unlike many traditional ML problems, these obstacles cannot be easily resolved simply by increasing data or computing power. Therefore, the necessity to continuously develop understandable and validatable methods in the future is underscored. This ongoing development is essential to ensure that power system protection maintains its robustness and adaptability to emerging challenges, as comprehensively described in Section 6. However, while these ongoing developments aim to tackle these challenges individually using different models, a universal ML model that addresses all the challenges highlighted in Section 5 is not yet available. Furthermore, to the best of our knowledge, a machine learning model has not been practically implemented in real-time protective relays by manufacturers, nor has any such experience been documented.

However, an analysis of the pie chart in Figure 5 reveals that a significant number of publications have focused on solving the problems of high impedance fault detection and voltage stability. This rise in research is credited to advancements in deep learning techniques, which have notably improved high impedance fault detection. Additionally, the deployment of Phasor Measurement Units in power networks has facilitated the development of machine learning models aimed at resolving voltage stability issues. Consequently, these two applications are likely to become a reality, potentially addressing the concerns discussed in Section 4. Despite the transformative potential of ML in various aspects of the power system, the continued reliance on proven, traditional methods for system protection is expected in the foreseeable future.

Author contributions

GP: Conceptualization, Writing—original draft, Writing—review & editing. KD: Writing—original draft, Writing—review & editing. JC: Writing—original draft, Writing—review & editing. DV: Writing—original draft, Writing—review & editing. AR: Project administration, Resources, Supervision, Writing—original draft, Writing—review & editing.

The author(s) declare that financial support was received for the research, authorship, and/or publication of this article. Partial financial support for this research project was received from Natural Sciences and Engineering Research Council of Canada: RGPIN-2018-05355 and Manitoba Hydro/Mitacs Inc. of Canada: IT28046. The funder was not involved in the study design, collection, analysis, interpretation of data, the writing of this article, or the decision to submit it for publication.

Conflict of interest

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

The author(s) declared that they were an editorial board member of Frontiers, at the time of submission. This had no impact on the peer review process and the final decision.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Aghababaeyan, Z., Abdellatif, M., Briand, L., Ramesh, S., and Bagherzadeh, M. (2023). Black-box testing of deep neural networks through test case diversity. IEEE Trans. Softw. Eng . 49, 3182–3204. doi: 10.1109/TSE.2023.3243522

Crossref Full Text | Google Scholar

Alam, M. K., Khan, F., Johnson, J., and Flicker, J. A. (2015). Comprehensive review of catastrophic faults in PV arrays: types, detection, and mitigation techniques. IEEE J. Photovolt . 5, 982–997. doi: 10.1109/JPHOTOV.2015.2397599

Alhelou, H. H., Hamedani-Golshan, M. E., Njenda, T. C., and Siano, P. A. (2019). survey on power system blackout and cascading events: research motivations and challenges. Energies 12, 1–28. doi: 10.3390/en12040682

Alimi, O. A., Ouahada, K., and Abu-Mahfouz, A. M. (2020). A review of machine learning approaches to power system security and stability. IEEE Access 8, 113512–113531. doi: 10.1109/ACCESS.2020.3003568

Aljohani, O., and Abu-Siada, A. (2016). Application of digital image processing to detect short-circuit turns in power transformers using frequency response analysis. IEEE Trans. Industr. Inform . 12, 2062–2073. doi: 10.1109/TII.2016.2594773

Amjady, N., and Majedi, S. F. (2007). Transient stability prediction by a hybrid intelligent system. IEEE Trans. Power Syst. 22, 1275–1283. doi: 10.1109/TPWRS.2007.901667

Azhar, I. F., Putranto, L. M., and Irnawan, R. (2022). Development of PMU-based transient stability detection methods using CNN-LSTM considering time series data measurement. Energies 15:21. doi: 10.3390/en15218241

Badrinarayanan, V., Kendall, A., and Cipolla, R. (2017). SegNet: a deep convolutional encoder-decoder architecture for image segmentation. IEEE Trans. Pattern Anal. Mach. Intell . 39, 2481–2495. doi: 10.1109/TPAMI.2016.2644615

PubMed Abstract | Crossref Full Text | Google Scholar

Banerjee, A., Kavasseri, R. G., Gani, M. B., and Brahma, S. (2019). “Towards supervisory protection using energy functions for relay misoperations in a stressed power system during blackout,” in 2019 IEEE Milan PowerTech . IEEE, 1–6.

Google Scholar

Behdadnia, T., Yaslan, Y., and Genc, I. A. (2021). new method of decision tree based transient stability assessment using hybrid simulation for real-time PMU measurements. IET Gen. Trans. Distrib . 15, 678–693. doi: 10.1049/gtd2.12051

Behkam, R., Karami, H., Naderi, M. S., and Gharehpetian, G. B. (2022a). “Condition monitoring of distribution transformers using frequency response traces and artificial neural network to detect the extent of windings axial displacements,” in 2022 26th International Electrical Power Distribution Conference (EPDC) . IEEE, 18–23.

Behkam, R., Karami, H., Naderi, M. S., and Gharehpetian, G. B. (2022b). “Intelligent interpretation of frequency response signatures to diagnose radial deformation in transformer windings using artificial neural network,” in 2022 12th International Conference on Computer and Knowledge Engineering (ICCKE) . IEEE, 523–528.

Behkam, R., Karami, H., Naderi, M. S., and Gharehpetian, G. B. (2022c). Application of artificial neural network on diagnosing location and extent of disk space variations in transformer windings using frequency response analysis,” in 2022 30th International Conference on Electrical Engineering (ICEE) . IEEE, 1079–1084.

Biagioni, D., Zhang, X., Wald, D., Vaidhynathan, D., Chintala, R., King, J., et al. (2022). “Powergridworld: A framework for multi-agent reinforcement learning in power systems,” in Proceedings of the Thirteenth ACM International Conference on Future Energy Systems , 565–570.

Bigdeli, M., Siano, P., and Alhelou, H. H. (2021). Intelligent classifiers in distinguishing transformer faults using frequency response analysis. IEEE Access 9, 13981–13991. doi: 10.1109/ACCESS.2021.3052144

Blackburn, J. L., and Domin, T. J. (2006). Protective Relaying: Principles and Applications . London: CRC Press, 695.

Bo, Q., Wang, X., and Liu, K. (2014). “Minimum frequency prediction of power system after disturbance based on the v-support vector regression,” in 2014 International Conference on Power System Technology . IEEE, 614–619.

Bragone, F., Morozovska, K., Hilber, P., Laneryd, T., and Luvisotto, M. (2022). Physics-informed neural networks for modelling power transformer's dynamic thermal behaviour. Electric Power Syst. Res . 211:108447. doi: 10.1016/j.epsr.2022.108447

Busoniu, L., Babuska, R., and De Schutter, B. (2008). A comprehensive survey of multiagent reinforcement learning. IEEE Trans. Syst. Man Cybernetics Appl. Rev. 38, 156–172. doi: 10.1109/TSMCC.2007.913919

Canese, L., Cardarilli, G. C., Di Nunzio, L., Fazzolari, R., Giardino, D., Re, M., et al. (2021). Multi-agent reinforcement learning: a review of challenges and applications. Appl. Sci. 11:4948. doi: 10.3390/app11114948

Cepeda, J., Gómez, I., Calero, F., and Vaca, A. (2022). “Big data platform for real-time oscillatory stability predictive assessment using recurrent neural networks and WA protector's records,” in 2022 International Conference on Smart Grid Synchronized Measurements and Analytics (SGSMA) . IEEE, 1–6.

Chellappa, R., Theodoridis, S., and Van Schaik, A. (2021). Advances in machine learning and deep neural networks. Proc. IEEE 109, 607–611. doi: 10.1109/JPROC.2021.3072172

Chen, C., Cui, M., Li, F., Yin, S., and Wang, X. (2021). Model-free emergency frequency control based on reinforcement learning. IEEE Trans. Industr. Inform 17, 2336–2346. doi: 10.1109/TII.2020.3001095

Chen, Y., Tan, Y., and Deka, D. (2018). “Is Machine Learning in Power Systems Vulnerable?,” 2018 IEEE International Conference on Communications, Control, and Computing Technologies for Smart Grids, SmartGridComm.

Chu, T., Chinchali, S., and Katti, S. (2020). Multi-agent reinforcement learning for networked system control.

CIGRE (2001). System Protection Schemes in Power Networks (Task Force 38, 02.19). Available online at: https://cigreindia.org/CIGRE%20Lib/Tech.%20Brochure/187%20System%20Protection%20schmes%20in%20power%20system.pdf (accessed February 19, 2001).

Cui, F., Cui, Q., and Song, Y. A. (2021). Survey on learning-based approaches for modeling and classification of human–machine dialog systems. IEEE Trans. Neural. Netw. Learn Syst . 32, 1418–1432. doi: 10.1109/TNNLS.2020.2985588

Dahal, O. P., Brahma, S. M., and Cao, H. (2014). Comprehensive clustering of disturbance events recorded by phasor measurement units. IEEE Trans. Power Deliv . 29, 1390–1397. doi: 10.1109/TPWRD.2013.2285097

Dharmapala, K., and Rajapakse, A. (2024). Short-term voltage instability prediction using pre-identified voltage templates and machine learning classifiers. Int. J. Electr. Power Energ. Syst . 156:109758. doi: 10.1016/j.ijepes.2023.109758

Dharmapala, K. D., Rajapakse, A., Narendra, K., and Zhang, Y. (2020). Machine learning based real-time monitoring of long-term voltage stability using voltage stability indices. IEEE Access 8, 222544–222555. doi: 10.1109/ACCESS.2020.3043935

Dorado-Rojas, S. A., Bogodorova, T., and Vanfretti, L. (2021). “Time series-based small-signal stability assessment using deep learning,” in 2021 North American Power Symposium (NAPS) . IEEE, 1–6.

Doshi-Velez, F., and Kim, B. (2017). Towards a rigorous science of interpretable machine learning. arXiv [Preprint]. arXiv:1702.08608.

Došilović, F. K., Brčić, M., and Hlupić, N. (2018). Explainable artificial intelligence: A survey,” in 2018 41st International Convention on Information and Communication Technology, Electronics and Microelectronics (MIPRO) . IEEE, 210–215. doi: 10.23919/MIPRO.2018.8400040

Duan, L., Hu, J., Zhao, G., Chen, K., Wang, S. X., He, J., et al. (2019). Method of inter-turn fault detection for next-generation smart transformers based on deep learning algorithm. High Voltage . 4, 282–291. doi: 10.1049/hve.2019.0067

Ernst, D., Glavic, M., and Wehenkel, L. (2004). Power systems stability control: reinforcement learning framework. IEEE Trans. Power Syst. 19, 427–435. doi: 10.1109/TPWRS.2003.821457

Eskandari, A., Milimonfared, J., and Aghaei, M. (2021). Fault detection and classification for photovoltaic systems based on hierarchical classification and machine learning technique. IEEE Trans. Ind. Electr . 68, 12750–12759. doi: 10.1109/TIE.2020.3047066

Esmaeilian, A., Popovic, T., and Kezunovic, M. (2015). Transmission line relay mis-operation detection based on time-synchronized field data. Electric Power Syst. Res . 125, 174–183. doi: 10.1016/j.epsr.2015.04.008

Fawcett, T. (2006). An introduction to ROC analysis. Pattern Recognit Lett . 27, 861–874. doi: 10.1016/j.patrec.2005.10.010

Gheisarnejad, M., and Khooban, M. H. (2020). IoT-Based DC/DC deep learning power converter control: real-time implementation. IEEE Trans. Power Electron . 35, 13621–13630. doi: 10.1109/TPEL.2020.2993635

Gomez, F. R., Rajapakse, A. D., Annakkage, U. D., and Fernando, I. T. (2011). Support vector machine-based algorithm for post-fault transient stability status prediction using synchronized measurements. IEEE Trans. Power Syst. 26, 1474–1483. doi: 10.1109/TPWRS.2010.2082575

Gonzalez, T. F. (2007). Handbook of Approximation Algorithms and Metaheuristics . London: Chapman and Hall/CRC.

Goodfellow, I. J., Shlens, J., and Szegedy, C. (2015). “Explaining and harnessing adversarial examples,” in 3rd International Conference on Learning Representations, ICLR 2015 - Conference Track Proceedings , 1–11.

Hadidi, R., and Jeyasurya, B. (2009). Reinforcement learning approach for controlling power system stabilizers. Can. J. Electr. Comput. Eng . 34, 99–103. doi: 10.1109/CJECE.2009.5443857

Hasani, R., Lechner, M., Amini, A., Rus, D., and Grosu, R. (2021). Liquid time-constant networks. Proc. AAAI Conf. AI 35, 7657–7666. doi: 10.1609/aaai.v35i9.16936

Hatcher, W. G., and Yu, W. A. (2018). Survey of deep learning: platforms, applications and emerging research trends. IEEE Access . 6, 24411–24432. doi: 10.1109/ACCESS.2018.2830661

Hatziargyriou, N., Milanovic, J., Rahmann, C., Ajjarapu, V., Canizares, C., Erlich, I. C., et al. (2020). Definition and classification of power system stability–revisited and extended. IEEE Trans. Power Syst. 36, 3271–3281. doi: 10.1109/TPWRS.2020.3041774

Hossain, E., Hossain, J., and Un-Noor, F. (2018). Utility grid: present challenges and their potential solutions. IEEE Access 6, 60294–60317. doi: 10.1109/ACCESS.2018.2873615

Hotelling, H. (1933). Analysis of a complex statistical variables into principal components 8. Determination of principal components for individuals. J. Educ. Psychol . 24, 498–520. doi: 10.1037/h0070888

Hou, Z. J., Ren, H., Wang, H., and Etingov, P. (2020). “Spatiotemporal pattern recognition in the PMU signals in the WECC system,” in 2020 IEEE Power and Energy Society General Meeting (PESGM) . IEEE, 1–5.

Huang, B., and Wang, J. (2022). Applications of physics-informed neural networks in power systems-a review. IEEE Trans. Power Syst. 38, 572–588. doi: 10.1109/TPWRS.2022.3162473

Huang, Q., Huang, R., Hao, W., Tan, J., Fan, R., Huang, Z., et al. (2020). Adaptive power system emergency control using deep reinforcement learning. IEEE Trans. Smart Grid . 11, 1171–1182. doi: 10.1109/TSG.2019.2933191

Huang, W. R., Emam, Z., Goldblum, M., Fowl, L., Terry, J. K, Huang, F., et al. (2020). Understanding generalization through visualizations. arXiv [Preprint]. arXiv:1906.03291.

IEEE (1988). IEEE Recommended Practice for the Design of Reliable Industrial and Commercial Power Systems. Piscataway, NJ: IEEE.

Islam, M. T., Karim Siddique, B. M. N., Rahman, S., and Jabid, T. (2018). Image recognition with deep learning. ICIIBMS 3, 106–110. doi: 10.1109/ICIIBMS.2018.8549986

Jayamaha, D. K. J. S., Lidula, N. W. A., and Rajapakse, A. D. (2019). “Wavelet based artificial neural networks for detection and classification of DC microgrid faults,” in 2019 IEEE Power and Energy Society General Meeting (PESGM) . IEEE, 1–5.

Jin, Z., Shang, J., Zhu, Q., Ling, C., Xie, W., Qiang, B., et al. (2020). “RFRSF: Employee turnover prediction based on random forests and survival analysis,” in Web Information Systems Engineering–WISE 2020, 21st International Conference, Amsterdam, The Netherlands, October 20–24, 2020, Proceedings, Part II 21 . Cham: Springer International Publishing, 503–515.

Kamwa, I., Samantaray, S. R., and Joós, G. (2009). Development of rule-based classifiers for rapid stability assessment of wide-area post-disturbance records. IEEE Trans. Power Syst. 24, 258–270. doi: 10.1109/TPWRS.2008.2009430

Karlsson, D., and Hill, D. J. (1994). Modelung and identification of nonlinear dynamic loads in power systems. IEEE Trans. Power Syst . 9, 157–166. doi: 10.1109/59.317546

Kaygusuz, C., Babun, L., Aksu, H., and Uluagac, A. S. (2018). “Detection of compromised smart grid devices with machine learning and convolution techniques,” in 2018 IEEE International Conference on Communications (ICC) . IEEE, 1–6.

Khalili Senobari, R., Sadeh, J., and Borsi, H. (2018). Frequency response analysis (FRA) of transformers as a tool for fault detection and location: a review. Electric Power Syst. Res . 155, 172–183. doi: 10.1016/j.epsr.2017.10.014

Khenak, F. (2010). Q-learning. CISIM 292, 228–232. doi: 10.1109/CISIM.2010.5643660

Kumar, U., Mishra, S., and Dash, K. (2023). An IoT and semi-supervised learning-based sensorless technique for panel level solar photovoltaic array fault diagnosis. IEEE Trans. Instrum. Meas . 72, 1–12. doi: 10.1109/TIM.2023.3287247

Lechner, M., Hasani, R., Amini, A., Henzinger, T. A., Rus, D., Grosu, R., et al. (2020). Neural circuit policies enabling auditable autonomy. Nat. Mach Intell . 2, 642–652. doi: 10.1038/s42256-020-00237-3

Lecun, Y., Bottou, L., Bengio, Y., and Ha, P. (1998). LeNet. Proc. IEEE 5, 1–46.

Li, J., Chen, S., Wang, X., and Pu, T. (2022). Load shedding control strategy in power grid emergency state based on deep reinforcement learning. CSEE J. Power Energ. Syst . 8, 1175–1182. doi: 10.17775/CSEEJPES.2020.06120

Li, Y., Gao, W., Huang, S., Wang, R., Yan, W., Gevorgian, V., et al. (2021). Data-driven optimal control strategy for virtual synchronous generator via deep reinforcement learning approach. J. Modern Power Syst. Clean Energ . 9, 919–929. doi: 10.35833/MPCE.2020.000267

Li, Z., Zhang, Y., Abu-Siada, A., Chen, X., Li, Z., Xu, Y., et al. (2021). Fault diagnosis of transformer windings based on decision tree and fully connected neural network. Energies 14, 1–6. doi: 10.3390/en14061531

Lidula, N. W. A., Perera, N., and Rajapakse, A. D. (2009). “Investigation of a fast islanding detection methodology using transient signals,” in 2009 IEEE Power and Energy Society General Meeting . IEEE, 1–6.

Lidula, N. W. A., and Rajapakse, A. D. A. (2010). Pattern recognition approach for detecting power islands using transient signals—part I: design and implementation. IEEE Trans. Power Deliv . 25, 3070–3077. doi: 10.1109/TPWRD.2010.2053724

Lidula, N. W. A., and Rajapakse, A. D. A. (2012). Pattern-recognition approach for detecting power islands using transient signals—part ii: performance evaluation. IEEE Trans. Power Deliv . 27, 1071–1080. doi: 10.1109/TPWRD.2012.2187344

Liu, J., Zhao, Z., Tang, C., Yao, C., Li, C., Islam, S., et al. (2019). Classifying transformer winding deformation fault types and degrees using FRA based on support vector machine. IEEE Access 7, 112494–112504. doi: 10.1109/ACCESS.2019.2932497

Liu, S., Mao, D., Xue, T., Tang, F., Li, X., Liu, L., et al. (2021). A data-driven approach for online inter-area oscillatory stability assessment of power systems based on random bits forest considering feature redundancy. Energies 14:6. doi: 10.3390/en14061641

Lu, S., Sirojan, T., Phung, B. T., Zhang, D., and Ambikairajah, E. (2019). An effective methodology for DC series arc fault diagnosis in photovoltaic systems. IEEE Access 7, 45831–45840. doi: 10.1109/ACCESS.2019.2909267

Machlev, R., Heistrene, L., Perl, M., Levy, K. Y., Belikov, J., Mannor, S., et al. (2022). Explainable artificial intelligence (XAI) techniques for energy and power systems: Review, challenges and opportunities. Energ. AI 9:100169. doi: 10.1016/j.egyai.2022.100169

Madeti, S. R., and Singh, S. N. (2018). Modeling of PV system based on experimental data for fault detection using kNN method. Solar Energ . 173, 139–151. doi: 10.1016/j.solener.2018.07.038

Mahadevkar, S. V., Khemani, B., Patil, S., Kotecha, K., Vora, D. R., Abraham, A., et al. (2022). A review on machine learning styles in computer vision—techniques and future directions. IEEE Access 10, 107293–107329. doi: 10.1109/ACCESS.2022.3209825

Makarov, Y. V., Lu, S., Samaan, N., Huang, Z., Subbarao, K., Etingov, P. V., et al. (2011). “Integration of uncertainty information into power system operations,” in 2011 IEEE Power and Energy Society General Meeting . IEEE, 1–13.

Marin-Quintero, J., Orozco-Henao, C., Bretas, A. S., Velez, J. C., Herrada, A., Barranco-Carlos, A., et al. (2022). Adaptive fault detection based on neural networks and multiple sampling points for distribution networks and microgrids. J. Modern Power Syst. Clean Energ . 10, 1648–1657. doi: 10.35833/MPCE.2021.000444

Mehta, P., Bukov, M., Wang, C. H., Day, A. G. R., Richardson, C., Fisher, C. K., et al. (2019). A high-bias, low-variance introduction to Machine Learning for physicists. Phys. Rep . 810, 1–124. doi: 10.1016/j.physrep.2019.03.001

Misyris, G. S., Venzke, A., and Chatzivasileiadis, S. (2020). “Physics-informed neural networks for power systems,” in 2020 IEEE Power and Energy Society General Meeting (PESGM ). IEEE, 1–5.

Mnih, V., Kavukcuoglu, K., Silver, D., Graves, A., Antonoglou, I., Wierstra, D., et al. (2013). Playing atari with deep reinforcement learning. arXiv [Preprint]. arXiv:1312.5602.

Molnar, C. (2020). Interpretable Machine Learning . London: Lulu. com.

Moravej, Z., Mortazavi, S. H., and Shahrtash, S. M. (2015). DT-CWT based event feature extraction for high impedance faults detection in distribution system. Int. Trans. Electr. Energy Syst. 25, 3288–3303. doi: 10.1002/etep.2035

MRO Protective Relay Subcommittee (2017). PRS Phase II Misoperations White Paper . Midwest Reliability Organization. Available online at: https://www.mro.net/document/protective-relay-subcommittee-misoperations%20phase-ii-whitepaper/?download

Nikoofekr, I., Sarlak, M., and Shahrtash, S. M. (2013). “Detection and classification of high impedance faults in power distribution networks using ART neural networks,” in 2013 21st Iranian Conference on Electrical Engineering (ICEE ). IEEE, 1–6.

Pan, S. J., and Yang, Q. (2010). A survey on transfer learning. IEEE Trans. Knowl. Data Eng . 22, 1345–1359. doi: 10.1109/TKDE.2009.191

Pani, S. R., Bera, P. K., and Kumar, V. (2020). “Detection and classification of internal faults in power transformers using tree based classifiers,” in 2020 IEEE International Conference on Power Electronics, Drives and Energy Systems (PEDES) . IEEE, 1–6.

Pedro, D. A. (2012). Few useful things to know about machine learning. Commun. ACM 55, 79–88. doi: 10.1145/2347736.2347755

Phadke, A., Wall, P., Ding, L., and Terzija, V. (2016). Improving the performance of power system protection using wide area monitoring systems. J. Modern Power Syst. Clean Energ . 4, 1–12. doi: 10.1007/s40565-016-0211-x

Qiu, J., Wu, Q., Ding, G., Xu, Y., and Feng, S. (2016). A survey of machine learning for big data processing. EURASIP J. Adv. Sig. Proc. 2016, 1–16. doi: 10.1186/s13634-016-0355-x

Rai, K., Hojatpanah, F., Badrkhani Ajaei, F., and Grolinger, K. (2021). Deep learning for high-impedance fault detection: convolutional autoencoders. Energies 14:623. doi: 10.3390/en14123623

Raissi, M., Perdikaris, P., and Karniadakis, G. E. (2019). Physics-informed neural networks: a deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. J. Comput. Phys . 378, 686–707. doi: 10.1016/j.jcp.2018.10.045

Rajapakse, A., Puangpairoj, A., Chirarattananon, S., and Thukaram, D. (2002). “Harmonic minimizing neural network SVC controller for compensating unbalanced fluctuating loads,” in 10th International Conference on Harmonics and Quality of Power. Proceedings , 403–408. IEEE.

Rajapakse, A. D., Gomez, F., Nanayakkara, K., Crossley, P. A., and Terzija, V. V. (2010). Rotor angle instability prediction using post-disturbance voltage trajectories. IEEE Trans. Power Syst. 25, 947–956. doi: 10.1109/TPWRS.2009.2036265

Ray, S. A. (2019). “Quick review of machine learning algorithms,” in Proceedings of the International Conference on Machine Learning, Big Data, Cloud and Parallel Computing: Trends, Prespectives and Prospects, COMITCon , 35–39.

Report, T. (2014). “Communication networks and systems for power utility automation - Part 1: Introduction and overview” in IEC TR 61850-1:2013 . 1–73.

Ribani, R., and Marengoni, M. (2019). “A survey of transfer learning for convolutional neural networks,” in 2019 32nd SIBGRAPI Conference on Graphics, Patterns and Images Tutorials (SIBGRAPI-T) . IEEE, 47–57.

Ribeiro, M. T., Singh, S., and Guestrin, C. (2016). “Why should i trust you?” Explaining the predictions of any classifier,” in Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , 1135–1144.

Rojas-Dueñas, G., Riba, J. R., Kahalerras, K., Moreno-Eguilaz, M., Kadechkar, A., Gomez-Pau, A., et al. (2020). “Black-box modelling of a dc-dc buck converter based on a recurrent neural network,” in 2020 IEEE International Conference on Industrial Technology (ICIT) . IEEE, 456–461.

Rudin, C., Waltz, D., Anderson, R., Boulanger, A., Salleb-Aouissi, A., Chow, M., et al. (2012). Machine learning for the New York City power grid. IEEE Trans. Pattern Anal. Mach. Intell . 34, 328–345. doi: 10.1109/TPAMI.2011.108

Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., et al. (2015). ImageNet large scale visual recognition challenge. Int. J. Comput. Vis . 115, 211–252. doi: 10.1007/s11263-015-0816-y

Safavian, S. R., and Landgrebe, D. A. (1991). Survey of decision tree classifier methodology. IEEE Trans. Syst. Man. Cybern . 21, 660–674. doi: 10.1109/21.97458

Sessions, V., and Valtorta, M. (2006). “The effects of data quality on machine learning algorithms,” in Proceedings of the 11th International Conference on Information Quality , eds J. R. Talburt, E. M. Pierce, N. Wu, and T. Campbell, Vol. 6 (Cambridge, MA: MIT), 485–498.

Silver, D., Hubert, T., Schrittwieser, J., Antonoglou, I., Lai, M., Guez, A., et al. (2017). Mastering chess and shogi by self-play with a general reinforcement learning algorithm. arXiv [Preprint]. arXiv:1712.01815.

PubMed Abstract | Google Scholar

Simeone, O. A. (2018). Very brief introduction to machine learning with applications to communication systems. IEEE Trans. Cogn. Commun. Netw . 4, 648–664. doi: 10.1109/TCCN.2018.2881442

Simões, L. D., De Oliveira, A. L. R., Bezerra Costa, F., and Prado De Medeiros, R. (2021). A Machine Learning-Based Internal Fault Identification in Power Transformers.

Sirojan, T., Lu, S., Phung, B. T., Zhang, D., and Ambikairajah, E. (2022). sustainable deep learning at grid edge for real-time high impedance fault detection. IEEE Trans. Sust. Comput . 7, 346–357. doi: 10.1109/TSUSC.2018.2879960

Srivastava, A., and Parida, S. K. A. (2022). Robust fault detection and location prediction module using support vector machine and gaussian process regression for AC microgrid. IEEE Trans. Ind. Appl . 58, 930–939. doi: 10.1109/TIA.2021.3129982

Subramanian, M. (2020). Detection of winding inter-turn faults. Transf. Magazine . 7:1. Available online at: https://transformers-magazine.com/magazine/7243-fault-localization-using-sfra-part-i/

Sutskever, I., Vinyals, O., and Le, Q. V. (2014). Sequence to sequence learning with neural networks. Adv. Neural. Inf. Process. Syst . 4, 3104–3112.

Sutton, R. S. (1988). Learning to predict by the methods of temporal differences. Mach. Learn . 3, 9–44. doi: 10.1007/BF00115009

Syed, D., Zainab, A., Ghrayeb, A., Refaat, S. S., Abu-Rub, H., Bouhali, O., et al. (2021). Smart grid big data analytics: survey of technologies, techniques, and applications. IEEE Access 9, 59564–59585. doi: 10.1109/ACCESS.2020.3041178

Teeuwsen, S. P., Erlich, I., El-Sharkawi, M. A., and Bachmann, U. (2006). Genetic algorithm and decision tree-based oscillatory stability assessment. IEEE Trans. Power Syst. 21, 746–753. doi: 10.1109/TPWRS.2006.873408

Thomas, A., Koshy, S., and Sunitha, R. (2020). “Machine learning based detection and classification of power system events,” in 2020 International Conference on Power, Instrumentation, Control and Computing (PICC) . IEEE, 1–6.

Tripathi, S. (2018). Dynamic prediction DS of powerline frequency for wide area monitoring and control. IEEE Trans. Industr. Inform 14, 2837–2846. doi: 10.1109/TII.2017.2777148

Ustun, T. S., Hussain, S. M. S., Yavuz, L., and Onen, A. (2021a). Artificial intelligence based intrusion detection system for IEC 61850 sampled values under symmetric and asymmetric faults. IEEE Access 9, 56486–56495. doi: 10.1109/ACCESS.2021.3071141

Ustun, T. S., Suhail Hussain, S. M., Ulutas, A., Onen, A., Roomi, M. M., Mashima, D., et al. (2021b). Machine learning-based intrusion detection for achieving cybersecurity in smart grids using IEC 61850 GOOSE messages. Symmetry 13:5. doi: 10.3390/sym13050826

Vega, T. Y., Roig, V. F., and San Segundo, H. B. (2007). “Evolution of signal processing techniques in power quality,” in 2007 9th International Conference on Electrical Power Quality and Utilisation . IEEE, 1–5.

Vosoughi, A., and Samimi, M. H. (2022). “Evaluation of the Image Processing Technique in Interpretation of Polar Plot Characteristics of Transformer Frequency Response,” in 2022 International Conference on Machine Vision and Image Processing (MVIP) . IEEE, 1–6.

Vu, T. L., Mukherjee, S., Yin, T., Huang, R., Tan, J., Huang, Q., et al. (2021). “Safe reinforcement learning for emergency load shedding of power systems,” in 2021 IEEE Power and Energy Society General Meeting (PESGM) . IEEE, 1–5.

Wang, J., Xu, W., Gu, Y., Song, W., and Green, T. C. (2021). Multi-agent reinforcement learning for active voltage control on power distribution networks. Adv. Neural Inf. Proc. Syst. 34, 3271–3284.

Western Electricity Coordinating Council (2018). Misoperations Reduction Strategy. Available online at: https://www.wecc.org/Reliability/Misoperations%20Reduction%20Strategy%20Review.pdf

Wischkaemper, J., and Brahma, S. (2021). Machine learning and power system protection. IEEE Electr. Mag . 9, 112–114. doi: 10.1109/MELE.2020.3047031

Wong, T. T., and Yang, N. Y. (2017). Dependency analysis of accuracy estimates in k-fold cross validation. IEEE Trans. Knowl. Data Eng . 29, 2417–2427. doi: 10.1109/TKDE.2017.2740926

Wu, M., and Chen, L. (2016). Image recognition based on deep learning. Proc. CAC 2016, 542–546. doi: 10.1109/CAC.2015.7382560

Xie, J., and Sun, W. A. (2021). Transfer and deep learning-based method for online frequency stability assessment and control. IEEE Access 9, 75712–75721. doi: 10.1109/ACCESS.2021.3082001

Yamashita, K., Li, J., Zhang, P., and Liu, C. C. (2009). “Analysis and control of major blackout events,” in 2009 IEEE/PES Power Systems Conference and Exposition , PSCE, 1–4.

Yang, H., Zhang, W., Chen, J., and Wang, L. (2018). PMU-based voltage stability prediction using least square support vector machine with online learning. Electric Power Syst. Res . 160, 234–242. doi: 10.1016/j.epsr.2018.02.018

Yi, Z., and Etemadi, A. H. (2017). Line-to-line fault detection for photovoltaic arrays based on multi-resolution signal decomposition and two-stage support vector machine. IEEE Trans. Ind. Electr . 64, 8546–8556. doi: 10.1109/TIE.2017.2703681

Yu, N., Shah, S., Johnson, R., Sherick, R., Hong, M., Loparo, K., et al. (2015). “Big data analytics in power distribution systems,” in 2015 IEEE Power and Energy Society Innovative Smart Grid Technologies Conference (ISGT) . IEEE, 1–5.

Zaker, B., Gharehpetian, G. B., Mirsalim, M., and Moaddabi, N. (2013). “PMU-based linear and nonlinear black-box modelling of power systems,” in 2013 21st Iranian Conference on Electrical Engineering (ICEE) . IEEE, 1–6.

Zhang, K., Zhang, J., Xu, P. D., Gao, T., and Gao, D. W. (2022). Explainable AI in deep reinforcement learning models for power system emergency control. IEEE Trans. Comput. Soc. Syst . 9, 419–427. doi: 10.1109/TCSS.2021.3096824

Zhang, M., Li, J., Li, Y., and Xu, R. (2021). Deep learning for short-term voltage stability assessment of power systems. IEEE Access 9, 29711–29718. doi: 10.1109/ACCESS.2021.3057659

Zhang, Y., Ilić, M. D., and Tonguz, O. K. (2011). Mitigating blackouts via smart relays: a machine learning approach. Proc. IEEE 99, 94–118. doi: 10.1109/JPROC.2010.2072970

Zhao, T., Wang, J., Lu, X., and Du, Y. (2022). Neural lyapunov control for power system transient stability: a deep learning-based approach. IEEE Trans. Power Syst. 37, 955–966. doi: 10.1109/TPWRS.2021.3102857

Zhao, Y., Ball, R., Mosesian, J., De Palma, J. F., and Lehman, B. (2015). Graph-based semi-supervised learning for fault detection and classification in solar photovoltaic arrays. IEEE Trans. Power Electron . 30, 2848–2858. doi: 10.1109/TPEL.2014.2364203

Zhou, D. Q., Annakkage, U. D., and Rajapakse, A. D. (2010). Online monitoring of voltage stability margin using an artificial neural network. IEEE Trans. Power Syst. 25, 1566–1574. doi: 10.1109/TPWRS.2009.2038059

Zhu, L., Hill, D. J., and Lu, C. (2021). Intelligent short-term voltage stability assessment via spatial attention rectified RNN learning. IEEE Trans. Industr. Inform 17, 7005–7016. doi: 10.1109/TII.2020.3041300

Zhu, L., Lu, C., Kamwa, I., and Zeng, H. (2020). Spatial-temporal feature learning in smart grids: a case study on short-term voltage stability assessment. IEEE Trans. Industr. Inform 16, 1470–1482. doi: 10.1109/TII.2018.2873605

Zhu, L., Lu, C., and Sun, Y. (2016). Time series shapelet classification based online short-term voltage stability assessment. IEEE Trans. Power Syst. 31, 1430–1439. doi: 10.1109/TPWRS.2015.2413895

Zhuang, F., Qi, Z., Duan, K., Xi, D., Zhu, Y., Zhu, H., et al. (2020). A comprehensive survey on transfer learning. Proc. IEEE 109, 43–76. doi: 10.1109/JPROC.2020.3004555

Nomenclature and abbreviations

Keywords: power system protection, power system stability, machine learning, emergency control, deep learning, reinforcement learning

Citation: Porawagamage G, Dharmapala K, Chaves JS, Villegas D and Rajapakse A (2024) A review of machine learning applications in power system protection and emergency control: opportunities, challenges, and future directions. Front. Smart Grids 3:1371153. doi: 10.3389/frsgr.2024.1371153

Received: 15 January 2024; Accepted: 21 March 2024; Published: 09 April 2024.

Reviewed by:

Copyright © 2024 Porawagamage, Dharmapala, Chaves, Villegas and Rajapakse. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY) . The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

*Correspondence: Athula Rajapakse, athula.rajapakse@umanitoba.ca

This article is part of the Research Topic

Horizons in Smart Grids

• Open access
• Published: 19 June 2016

Developments of power system protection and control

• Z. Q. Bo 1 ,
• X. N. Lin 2 ,
• Q. P. Wang 1 ,
• Y. H. Yi 1 &
• F. Q. Zhou 1  

Protection and Control of Modern Power Systems volume  1 , Article number:  7 ( 2016 ) Cite this article

28k Accesses

55 Citations

Metrics details

Synchronized wide area communication has become a mature technology, which makes the real-time interaction between the substations and the wide area protection and control system possible. However, the present protection and control system to handle this real-time data has been recognized to be deficient. This paper begins by reviewing the development history of power system protection, with special attention paid to the recent development in the field of wide-area and integrated protections, in order to look into the future development of protection and control systems. Then the concept of integrated wide area protection and control is introduced, where it can be shown that a hierarchical protection and control system provides the protection and control for wide area or regional power substations/plants and their associated power networks. The system is mainly divided into three levels: the local, the substation/plant, and the wide area/regional. The integrated functions at each level are described in details with an aim to develop an optimal coordination mechanism between each level. The key element in the proposed system is the wide area real-time protection and control information platform, which not only enables the merger of three lines of defence for power system protection and control, but also provides a perfect tool for the application of cloud computing in substations and power networks.

Introduction

Power system protection emerged at the beginning of the last century, with the application of the first electro-mechanical overcurrent relay. The majority of the protection principles currently employed in protection relays were developed within the first three decades of the last century, such as overcurrent, directional, distance and differential protection, as shown in Fig.  1 . The development of modern science and technology, especially electronic and computer technology, promoted the development of relay technology, such as materials, components and the manufacturing process of the hardware structure of relay protection device. At the same time, great theoretical progress had been made in the relay protection software, algorithms, etc. As shown in Fig.  1 , the progress in modern technology stimulates the development in power system protection. In the last century from the emergence of protection to the end of the 1990s, the relay protection had gone through a number of development stages, migrating from electro-mechanical to semiconductor, and subsequently to integrated circuit and microprocessor technologies. Today, microprocessor-based digital and numeric relays are replacing conventional relays in all areas of power system protection. However, many of the same relaying principles of protection are still playing a dominant role to date. In the late 1960s, the application of a centralized substation protection system based on a centralized computer system was proposed [ 1 ]. This constitutes an important milestone in the history of power system protection. The idea fits well with the concept of an overall integrated protection where the protection package would not only oversee individual units of a plant but also a section of the network. However, the idea has not been widely applied until recently, since there were no available computer hardware/software or communication technologies to support such an idea. Since then, relay technology has enjoyed successful developments based on the application of digital techniques. The introduction of microprocessors into protection in the 1980s generally followed the conventional approach with the implementation of distributed processing platforms that concentrated on protecting individual units of the system. Limited integrated protection was provided in the form of back-up protection and thus remained a secondary function.

History of power system protection

Developments in the 1980s and 1990s provided new means to advance power system protection, especially the ‘Adaptive Protection’ and the Artificial Intelligence (AI)-based protection techniques proposed in the 1980s and 1990s. The adaptive protection started with the application of Inverse Definite Minimum Time Overcurrent (IDMT) protection in the early time of protection history. The concept played an important role in the 1980s with the progress of computing technology and associated control theory. It can be defined as a new type of relay protection which can change the performance, characteristics or set value according to the operation mode and fault condition of the power system. The basic idea of adaptive relay protection is to protect the power system as much as possible to improve the performance of the protection. Adaptive relay protection has the advantages of improving the response of the system, enhancing the reliability and improving the economic benefits. It has a wide application prospect in the field of distance protection, transformer protection, generator protection, autoreclosure and so on. Research has discovered that, to achieve the protection of the system adaptive to the operation mode and fault status, more detailed system operation and fault information are required through communication network.

The 1990s witnessed the rapid development of electronic and computer technology, the artificial intelligence technology such as artificial neural networks, genetic algorithms, evolutionary algorithm, fuzzy logic and other research applications, which have been applied to the field of relay protection. For example, artificial neural network (ANN) is used to achieve fault type identification, fault distance measurement, direction protection, and so on. Artificial neural network has the characteristics of distributed storage, parallel processing, self-organization and self-learning. The application of artificial intelligence will improve the speed and accuracy of fault detection and analysis, which represents the future development of an intelligent diagnosis system.

As a result of these developments, the performance of the protection relays has been improved. However, these developments have concentrated on the improvement of conventional relaying techniques, and no significant new relaying principles have been derived from the application of the Adaptive and AI techniques. At the same time in the 1990s, with the continuous expansion of the power network, the demand for fast fault clearance to improve system stability encouraged research into non-power system frequency fault detection techniques to increase the speed of the relay response. This led to the development of the so-called ‘transient based protection’ relays based on travelling wave and superimposed components, which utilizes the fault generated transients for transmission system protection. Studies have found that the fault generated high frequency transients can be detected and quantified, creating the possibility for developing new protection principles and techniques [ 2 ]. Considerable effort has now been devoted to research on high frequency transient detection. Another important milestone is the application of novel communication technique, the utilization of global positioning system (GPS) in power system protection as shown in Fig.  1 . In this respect, a number of new techniques has been proposed. In particular, the new proposed protection relay principle is able to provide protection for wide area power network [ 3 , 4 ]. Following the development, the concept of wide area protection focusing on control aspect has been presented [ 5 ].

In recent years, the dramatic growth in signal processing capability of relay platforms, and the availability of suitable communications schemes, have provided a new opportunity to revisit the concept of centralized protection. Research [ 6 ] on the concept of ‘Integrated Protection’ shows that information obtained from multiple power plants and components can be used to derive new protection principles and schemes, which could have significant advantages over the existing protection techniques based on the individual plant or component. In this respect, the substation area protection [ 7 ] has quickly become a practical development and application area. Furthermore, the development of information technology has resulted in the interests in utilizing cloud computing [ 8 ] and big data techniques, as shown in Fig.  1 , to improve the performance of power system protection and control.

On the other front, the developments in the field of substation integrated automation technology provide the technical basis for optimizing the combination and system integration of monitoring, control, protection and measurement device and system. The implementation of relay protection and integrated automation reflects in the integration and resource sharing, remote control and information sharing. Taking the remote terminal unit and microcomputer protection device as the core, the control, signal, measurement, billing and other circuits are integrated into the computer system to replace the traditional control protection cabinet, which can reduce the area and equipment investment and improve the reliability of the secondary system. With the advanced computer and communication network, the relay protection device is actually a high performance and multi-function computer, which is an intelligent terminal of the whole power system computer network. It can obtain any information about the operation and fault of the power system from the network, and can transmit any information of the protected components to the network control center or any terminal. Therefore, each microcomputer protection device can complete not only the relay protection function, but also the measurement, control, data communication and other functions under normal operation condition, and can also realize the integration of protection, control, measurement and data communication.

Recent development in power system protection and control

With the fast progress in high-speed communication network and information technology, there were significant developments in power system protection, power system control and wide area control in recent years, particularly in the wide-area and integrated protection.

Recent development

• Wide area protection

In recent years, the fast development in communication technologies makes the wide-area information exchange possible. In this respect, the emergence of the wide area measurement system provides a new idea for the design of power system protection systems. The first wide-area protection principle is derived from the transient based protection in 1996 [ 2 ], in which GPS time synchronization played a major role in the design [ 3 ]. This was immediately followed by a summary paper in 1997, which systematically outlines the concept of the so-called “wide area protection” [ 4 ], focusing principally on the control aspect of the area. The wide area protection based on novel algorithms, which is derived from the measurements of multiple information points, is able to provide fast, reliable and accurate fault clearance, analyse the effects on the system stability based on the fault system analysis and take necessary control measures to perform the functions of relay protection, security, and stability control in order to prevent voltage collapse. Wide area relay protection has quickly become a hot research topic with many research results published particularly in recent years.

Integrated protection

With the development of digital technology, more and more protection functions for any given apparatus (line, transformer, generator, etc.) have been implemented within one protective device to achieve a certain degree of integration. For example, a numeric line protection relay may have distance or current differential function as the main protection, and directional and overcurrent functions as the backup protection. The recent developments in microprocessor and communication techniques provided new means to derive new protection principles and schemes based on the information obtained from multiple power plants and components, which could have significant advantages over the existing protection techniques based on the individual plant or component [ 5 ]. Unlike centralized protection (or substation area protection), the integrated protection does not simply centralize the relay hardware/software, but concentrates on the developments of new concepts and algorithms based on multiple points of measurements; via this means it is hoped that the performance of protection can be improved significantly. There was also research in the field of integrated wide area protection [ 9 ].

Wide area control

The increased deployment of wide-area measurements will significantly enhance the power system wide-area power system operation and control. They provide voltage and current phasor information, synchronized with high precision to a common time reference provided by GPS. Therefore, a wide range of power system monitoring and control applications can be implemented in the system for improving system awareness and reliability, which includes enhanced state estimation based on mixed Remote Terminal Unit (RTU) and Phasor Measurement Unit (PMU) measurements [ 10 ], dynamic model online estimation and validation [ 11 ], real-time congestion management, real-time stability estimation [ 12 ], detection and damping of inter-area oscillations [ 13 ]. However, the most important and challengeable applications are the implementations of wide area stability real-time detection and control to prevent blackouts [ 14 ]. There was also research in the integrated protection and control [ 15 ].

New concept and development

Based on the developments mentioned above, a new concept of the integrated wide area protection and control (IWAPC) has been proposed recently. The main focus of the concept [ 16 , 17 ] is the integration between the protection and control, particularly at the wide-area or regional level, aimed to provide a number of benefits to the future protection and control system, e.g., the potential to merge the three lines of defence system and on-line self-healing decision making, in order to prevent cascading tripping of large area power network. The concept of integrated wide area protection and control is introduced, in which a three-level hierarchically coordinated system, supported by the specially designed real-time synchronised wide-area communication network, provides the protection and control for wide area or regional power substations/plants and their associated power network. The key element in the system is the integrated wide area protection and control information platform, which receives real-time synchronised data from the communication network to support the integration of protection and control at the wide area/regional level.

The information platform also supports the application of a cloud computing system, which is specially designed to implement a number of secondary functions for substations and power networks. In addition to the basic functions of relay protection, the platform should have a large capacity of fault information and data storage, fast data processing functions, powerful communication functions, and other protection, control devices and scheduling network to share the whole system data, information and network resources, and can also carry out remote monitoring with the computer monitoring system of substation communication. With the proposed platform, the architecture of future substation equipment may be reshaped to provide a flexible framework for building an interactive grid and subsequently improve the reliability and security of power grids.

Integrated wide area/regional protection and control

Architecture of integrated wide area protection and control.

The proposed integrated wide area or regional protection and control system (IWAPC) is illustrated in Fig.  2 . There have been fast developments in both power transmission and distribution networks, e.g., the series compensation in AC lines and high-voltage DC lines in transmission systems, distributed generation and energy storage in distribution systems, etc. These new developments result in far more complicated characteristics than that of conventional systems. Consequently, the existing protection and control system will no longer be effective to cope with the new systems, and this has led to the proposed IWAPC system. As is shown, the IWAPC system consists of different equipment at different layers: from bottom to top, there is the integrated multiple-function intelligent equipment at the local level; the substation communication network and the integrated substation protection and control at the substation level; the wide area communication network, the integrated wide area information platform and the integrated wide area (regional) protection and control at wide area level. The key parts of the system are the high-speed wide area communication network and the real-time synchronisation information platform.

Integrated wide area protection and control

The IWAPC is further extended to dispatching in order to achieve the integration of dispatching automation, protection and control of power grid, and according to the three-level dispatching (country, province, regional) architecture to implement the functions of regional protection, control and dispatching managements.

Multiple functions intelligent equipment at the local level

As shown in Fig.  2 , the Intelligent equipment at local level is an integrated multiple function secondary equipment in the substation, which mainly consists of the MU, intelligent terminal, metrology measurement, PMU & local protection. The equipment is responsible for sampling all real-time data and sending information to the integrated substation P&C and wide area P&C. It also receives and carries out the control commands from the integrated substation P&C and the IWAPC. The equipment can be integrated into primary power apparatuses and achieve local protection for 90 % of its associated line sections. It has a redundant configuration to ensure reliability, together with other integrated functions such as fault recorder, data storage and network analysis, etc.

Integrated substation protection and control at the substation/plant level

The substation P&C integrates functions of line, bus, transformer protections, switch failure; autoreclosure, automatic bus transfer, UFLS, UVLS, overload inter-tripping and substation control function, etc. It utilizes information from the entire substation to achieve substation backup protection and safety automatic control, etc. The CBs are used as units to configure the adaptive backup protection, and current differential protection is used to replace the stage overcurrent protection, breaker failure protection and dead zone protection in the conventional protection system.

The IWAPC specially designed for the protection and control of power network is able to offer fast protection. In addition, they both integrate functions of automatic UFLS and UVLS, voltage and frequency control, oscillation detection and out-of-step separation, etc. In addition, the IWAPC also incorporates the function of transmission cross-section safety P&C. Unlike conventional protection and control, which are separated in both design and operation, the IWAPC integrates protection and control into one optimal combined system, which effectively coordinates the wide area (regional) protection and control, in order to achieve significant improvements in the protection and control of power systems.

Synchronised high speed communication network

One of the most important elements of the IWAPC system is the fast communication network. In this respect, the latest development in communication network, the Packet Transport Network (PTN) may be a better choice to implement such a task. The present power communication network is mainly used in multi-service transport platform based on the Synchronous Digital Hierarchy (SDH). Its advantages lie in its high efficiency for carrying TDM services, low latency, high reliability, with end management capabilities. However, with the new trends in smart grid development, SDH technology gradually revealed its limitations, such as low bearing efficiency and poor flexibility for data services. In contrast, PTN can realise statistical multiplexing and efficient transfer of packet service by using packet-switched core, which can overcome the weaknesses of SDH rigid bandwidth. In addition, it can provide good quality of service, operation, administration and maintenance. Self-healing fibre optical network is employed to connect a number of substations in the region, to ensure full sharing of dynamic and transient information for all electrical measurements, breaker status and protection operations; using high reliability IEEE-1588 technology to ensure the synchronization timing of the sharing data, to prove the data for the integrated wide area protection and control. However, SDH is still an option for the task since it has been widely applied in power network.

Synchronized information platform

Substation is installed with a wide range of electrical equipment with complex designs and is difficult to maintain. With the continuing improvement in power system automation and the intelligence level, the system network has been expanding, along with the huge amount of information in protection and control. As each piece of information is collected and stored by different devices in each separate system, the interoperability of the internal power system data between systems is poor, whereas complex communication protocols tend to create information islands. Consequently, the measurement data and protection control mechanism cannot be shared, which restricts the information integration. The protection and control of smart grid requires dealing with the new situation demands of the application, in order to improve further the information platform capabilities for the future development of key technologies, and to make the information platform system more open.

The real-time synchronized information platform accurately collects wide area information and conducts data mining to investigate the logic relation between the real-time information to increase the sensitivity, reliability and fault tolerance capability. The data received from the platform includes static, dynamic, transient measurements and states of circuit breakers, etc. Valuable information is extracted from the data and allocated to various specially-designed computation algorithms in the platform to perform advanced functions of protection and control for the power network. In the platform, sets of data need to be transferred and their transferring speed depends on the application, e.g., slow speed for contingency analysis, near real time speed for monitoring, real time speed for control, and high speed for wide area protection,; in particular, time synchronization. The information can also include other types of data, such as the oil and ambient temperature of the transformer, wind speed and direction, sun intensity, etc. On the other hand, the information is stored in a hierarchical manner instead of a centralized one, which comprises the hierarchical protection and control system. Equipped with the latest high-speed synchronised communication technology, integrated with the advanced protection techniques and the latest developments in control system, the system offers not only fast protection, but also complete control of entire power network.

The advanced computing technology is introduced to establish a synchronized information platform for wide area protection and control, to build a panoramic operation and maintenance data collection network, providing a standardized interface to the terminal device, to form a resource sharing, flexible and interactive, open and ordered information platform. In summary, advanced computing technologies are used to build a distributed collaborative intelligent information platform, simplifying terminal data collection equipment, and breaking the barriers between protection and control systems at different substations through the specially designed synchronized information platform.

Wide area power cloud

Based on the information platform mentioned above, a distributed cloud system is designed to implement functions at substation and regional levels, such as wide area fault location, fault line selection, power quality monitoring, protection settings, etc. The extended functions also include the equipment monitoring, life cycle and operation management, as shown in Fig.  3 .

Structure of distributed power cloud

Currently, many kinds of secondary equipments achieving different functions are installed in each substation, and an increasing number of distributed energy resources of small capacity added to the system greatly increase the number of equipments. To implement these equipments, complex functions in a specially developed distributed “cloud” system will greatly reduce the equipment investment. The cloud at substation level receives the data from process level, and the regional cloud receives the data from the information platform, which includes static, dynamic, transient measurements and states of circuit breakers, extracting valuable information and allocating them to various specially-designed computation algorithms in the platform to perform advanced functions in order to identify the faulted line, the accurate fault location and the contents of harmonics, etc.

The cloud computing platform can make full use of “processing ability of cloud” to reduce the burden of terminal secondary equipment. Based on big data technique, the computing clouds enjoy strong processing power based on demand. There is no need for endless upgrades to improve the processing capacity of the equipment, and there is also no need to update the software to achieve a variety of task processing. There are many more advantages which can be derived from the cloud system, such as the wide area information sharing, standardization of software and algorithm, reduction of equipment investment, substation area occupation and work load for operation and maintenance.

Conclusions

This paper presents an integrated wide area protection and control system based on a hierarchical structure, which integrates protection and control at local, substation and regional levels. Covering both transmission and distribution networks, the system is supported by the proposed high-speed synchronised communication network and the real-time protection and control information platform. The system, which integrates the advanced protection techniques and the latest developments in control system, offers not only fast protection, but also complete control of the entire power network. It offers a potential for the merger of the three lines of defence into a unified system to ensure more effectively the reliable and safe operation of power grid. Based on the system information platform, a distributed power cloud system is also designed to support many advanced applications for the integrated wide area protection and control.

With the continuous advances in measurement, communication and information technologies, the system presents a bright future for practical application. Overall improved performance of protection and control can be expected from the proposed system. However, for the system to become useful in power system application, it is equally important that its practical implementation be readily manageable, user-friendly and cost-effective. The authors hold that, to achieve these goals, the proposed integrated protection and control at wide area level offers an appealing way forward.

Rockefeller, G. D. (1969). Fault protection with digital computer. IEEE Transactions on Power Apparatus and Systems, 88 (4), 438–461.

Article   Google Scholar  

Bo, Z.Q., Jiang, F., Chen, Z., et al. (2000). Transient based protection for power transmission systems. In: IEEE PES Winter Meeting . Singapore.

Bo, Z.Q., Jayasinge, J.A.S.B., Aggarwal, R.K., et al. (1996). A new scheme for monitoring and protection of power transmission system based on global positioning system. In: Technol. Educ. Inst. Iraklio The 31st University Power Engineering Conference (pp. 21–24). Crete.

Bo, Z. Q., Weller, G., & Lomas, T. (2000). Positional protection of transmission system using global positioning system. IEEE Transactions on Power Delivery, 15 (4), 1163–1168.

Ingelsson, B., Lindstrom, P. O., Karlsson, D., et al. (1997). Wide-area protection against voltage collapse. IEEE Computer Applications in Power, 10 (4), 30–35.

Bo, Z. Q., He, J. H., & Dong, X. Z. (2005). Integrated protection of power network. Relay, 33 (14), 33–41.

Google Scholar  

Gao, H. L., Liu, Y. Q., Su, J. J., et al. (2012). New type of substation-area backup protection for intelligent substation. In 2012 China Smart Grid Seminar .

Bo, Z.Q., Wang, L., Zhou, F.Q., et al. (2014). Substation cloud computing for secondary auxiliary equipment.In: IEEE Powercon2014 . Chengdu.

Bo, Z. Q., Zhang, B. H., Dong, X. Z., He, J. H., et al. (2013). The development of protection intellectuation and smart relay network. Power System Protection and Control, 41 (2), 1–12.

Guo, Y., Wu, W. C., Zhang, B. M., et al. (2015). A distributed state estimation method for power systems incorporated with linear and nonlinear models. International Journal of Electrical Power & Energy Systems, 64 , 608–616.

Chen, R.Z., Wu, W.C., Sun, H.B., Zhang, & B.M. (2013). A two-level online parameter identification approach. In: IEEE Power & Energy Society General Meeting . Vancouver.

Liu, M.C., Zhang, B.M., Yao, L.Z., et al. (2008). PMU based voltage stability analysis for the transmission corridor. In: IEEE The 3rd Int. Conf. on Electric Utility Deregulation and Restructing and Power Technologies . Nanjing.

He, J. B., Lu, C., Wu, X. C., et al. (2007). Design and experiment of heuristic adaptive HVDC supplementary damping controller based on online Prony analysis. In Proc. IEEE Power Engineering Society General Meeting .

Wu, W.C., Zhang, B.M., Sun, H.B., et al. (2010). Development and application of on-line dynamic security early warning and preventive control system in China. In: IEEE Power & Energy Society General Meeting . Mineapolis.

Wang, B., Dong, X. Z., Xu, F., Cao, R. B., Liu, K., & Bo, Z. Q. (2011). Analysis of data sharing for protection and control system in smart distribution substation”. In Proceedings of the CSEE (Vol. 31).

Bo, Z.Q., Ge, S.M., Wang, Q.P., Wang, L., Zhou, F.Q., Liu, X., & Fan, Z.F. (2015). Integrated wide area protection and control systems based on advanced communication network. In: ATLANTIS PRESS 2015 International Conference on Mechanical Engineering, Materials and Energy . Tianjin.

Bo, Z.Q., Wang, Q.P., Wang, L., Zhou, F.Q., Ge, S.M., & Zhang, B.M. “Architecture design for integrated wide area protection and control systems”, APPEEC, April 14 2015, The 7 th Asia-Pacific Power and Energy Engineering Conference (APPEEC 2015).

Download references

Authors’ contributions

ZB investigated the framework of integrated wide area protection and control, and drafted the manuscript. XL summarized the history of power system protection. QW summarized recent development of power system protection. YY summarized the wide area power cloud. FZ participated in typesetting and revision of the manuscript. All authors read and approved the final manuscript.

Competing interests

The authors declare that they have no competing interests.

About the authors

Zhiqian Bo received his BSc degree from the Northeastern University, China in 1982 and PhD degree from The Queen's University of Belfast, UK in 1988 respectively. From 1989 to 1997, he was with the Power and Energy Group, University of Bath, UK. He has been with ALSTOM Grid Automation and responsible for new technology development and international research collaboration from 1998 to 2012. Recently he joined the XUJI Group of the State Grid Corporation of China (SGCC) as the Chief Expert on smart grid.

Xiangning Lin (SM’08) received the M.S. and Ph.D. degree in school of electrical and electronic engineering of Huazhong University of Science and Technology (HUST), Wuhan, China. Currently, he is a professor at HUST. His research interests are modern signal processing and its application in power systems as well as power system protective relaying and control.

Qingping Wang was born in Hebei Province, China, in 1975. He received his Ph.D in 2002 from the Department of Electrical Engineering, Tianjin University, China. At present, he has been employed as CTO of XJ Group Corporate Research. His research interests include protection, substation & distribution automation, power system simulation and micro-grid. He is author or co-author of more than 40 journal papers and owner of more than 30 invention patents.

Yonghui Yi , born in 1969, PhD in electrical engineering. Currently he works as the Managing Director of Protection and Automation Company of XUJI Group, His main research areas are smart grid and power system protection.

Fengquan Zhou , born in 1969, is the director of R&D Center, XJ Group, obtained his first Degree and graduate Degree from Tsinghua University, finished his PHD research in McGill University in Canada. His research areas include energy efficiency, power system automation protection, distributed generation, renewable energy, smart grid and other fields.

Author information

Authors and affiliations.

XUJI Group Corporation, Xuchang, China

Z. Q. Bo, Q. P. Wang, Y. H. Yi & F. Q. Zhou

Huazhong University of Science and Technology, Wuhan, China

You can also search for this author in PubMed   Google Scholar

Corresponding author

Correspondence to Z. Q. Bo .

Rights and permissions

Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License ( http://creativecommons.org/licenses/by/4.0/ ), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.

Reprints and permissions

About this article

Cite this article.

Bo, Z.Q., Lin, X.N., Wang, Q.P. et al. Developments of power system protection and control. Prot Control Mod Power Syst 1 , 7 (2016). https://doi.org/10.1186/s41601-016-0012-2

Download citation

Received : 09 May 2016

Accepted : 09 May 2016

Published : 19 June 2016

DOI : https://doi.org/10.1186/s41601-016-0012-2

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Power system protection
• Integrated protection and control
• Information platform

The Protection of Power System Based on Artificial Intelligence

• Conference paper
• First Online: 08 August 2017
• Cite this conference paper

• Zewei Zhu 16  

Part of the book series: Advances in Intelligent Systems and Computing ((AISC,volume 613))

Included in the following conference series:

• First International Conference on Real Time Intelligent Systems

1777 Accesses

1 Citations

With the development of modern electric power system, high voltage, long distance, and large capacity transmission, national networking have become inevitable. Therefore, the higher requirements have been put forward to ensure the safety and stable operation of the power grid relay protection technology. The bus-bar protection with high reliability, good selectivity and fast action speed has been concerned with the relay protection workers. At present, more and more attention is paid to the relay protection system based on artificial intelligence technology. Based on the analysis of the shortcomings of the traditional bus relay protection, the adaptive and self-learning ability of artificial intelligence technology was used in this paper, the concept of bus relay protection based on artificial intelligence technology was put forward, the related protection model was established, and the simulation experiment was carried out to verify the feasibility of the protection. The model simulation experiments prove that the new model has the advantages of dynamic velocity across the board, which can not only automatically identify the various types of applause, but also can make full use of the system in the real-time information, and better realize the protection of power systems.

This is a preview of subscription content, log in via an institution to check access.

Access this chapter

• Available as PDF
• Read on any device
• Instant download
• Own it forever
• Available as EPUB and PDF
• Compact, lightweight edition
• Dispatched in 3 to 5 business days
• Free shipping worldwide - see info

Tax calculation will be finalised at checkout

Purchases are for personal use only

Institutional subscriptions

Zhang, H.: Current protection based on neural network. J. Guangdong Univ. Technol. 19 (3), 17–21 (2001)

Google Scholar  

Bing, O.: Principle and implementation of adaptive current instantaneous fault protection. In: China’s Institutions of Higher Learning Power System and Its Automation Major 18th Annual Conference Proceedings, Wuhan, pp. 834–839 (2002)

Coury, D.V., Thorp, J.S., Hopkinson, K.M., et al.: Agent technology applied to adaptive relay setting for multi-terminal lines. In: IEEE Power Engineering Society Summer Meeting, Seattle, USA, vol. 2, pp. 2296–2101 (2000)

Zhong, G., Bo, Z.: No channel transmission line adaptive protection fault analysis and protection principle. Autom. Electr. Power Syst. 26 (7), 33–37 (2003)

Chen, Y.: Relay protection system based on multi agent technology. Autom. Electr. Power Syst. 26 (12), 1–6 (2002)

Wuyungaowa: The development of artificial neural network. Fujian Comput. 17 , 12–15 (2004)

Yu, Y., Zheng, S., et al.: Based on neural network of transmission lines without channel protection. Electr. Power Sci. Eng. 2 , 9–11 (2004)

Kasztenny, B., Kuras, K.: A new algorithm for low-impedance of busbars. In: Power Engineering Society Summer Meeting, 2001, vol. 1, pp. 97–102. IEEE (2001)

Li, G., Teng, L., Liu, W., Zhen, T.: The application of wavelet transform in the microcomputer bus protection. Electr. Power Autom. Equip. 22 (08), 67–69 (2002)

Xin, L., Tang, G.: Bus power system microcomputer protection technology and its application. Comput. Appl. 20 (05), 48–50 (2000)

Hughes, R., Legrand, E.: Numerical busbar protection benefits of numerical technology in electrical substation. In: 2001 Seventh International Conference on (IEE) Developments in Power System Protection, vol. 9, pp. 463–466 (2001)

Chen, Y., Ying, X., Zhang, Z.: Based on the technology of relay protection system. Autom. Electr. Power Syst. 26 (12), 1–6 (2002)

Xu, S., Fang, X.: Restore the application of expert system in power system structure. Electr. Power Autom. Equip. 23 (5), 33–35 (2003)

Zhou, D., Guan, X., Sun, J.: Power system short-term load forecasting based on neural network research. Grid Technol. 26 (2), 10–14 (2002)

Wang, J., Guangyi, Z.: Neural network technology in the application of the synchronous generator excitation system. Comput. Appl. 23 (6), 24–25 (2003)

Download references

Author information

Authors and affiliations.

China Three Gorges University, Yichang, Hubei, China

You can also search for this author in PubMed   Google Scholar

Corresponding author

Correspondence to Zewei Zhu .

Editor information

Editors and affiliations.

Institute of Mathematics and Computer Science, Opole University , Opole, Opolskie, Poland

Jolanta Mizera-Pietraszko

Digital Information Research Foundation , Chennai, Tamil Nadu, India

Pit Pichappan

Rights and permissions

Reprints and permissions

Copyright information

© 2018 Springer International Publishing AG

About this paper

Cite this paper.

Zhu, Z. (2018). The Protection of Power System Based on Artificial Intelligence. In: Mizera-Pietraszko, J., Pichappan, P. (eds) Lecture Notes in Real-Time Intelligent Systems. RTIS 2016. Advances in Intelligent Systems and Computing, vol 613. Springer, Cham. https://doi.org/10.1007/978-3-319-60744-3_45

Download citation

DOI : https://doi.org/10.1007/978-3-319-60744-3_45

Published : 08 August 2017

Publisher Name : Springer, Cham

Print ISBN : 978-3-319-60743-6

Online ISBN : 978-3-319-60744-3

eBook Packages : Engineering Engineering (R0)

Share this paper

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Publish with us

Policies and ethics

• Find a journal
• Track your research

• Search Menu
• Sign in through your institution
• Advanced Articles
• Editor's Choice
• Author Guidelines
• Publish with us
• Submission Site
• Open Access
• Self-Archiving Policy
• About Clean Energy
• About the National Institute of Clean and Low-Carbon Energy
• Editorial Board
• Instructions for Reviewers
• Advertising & Corporate Services
• Journals Career Network
• Journals on Oxford Academic
• Books on Oxford Academic

Article Contents

Introduction, 1 methods and challenges addressed, 2 ai techniques, 3 application of ai in the power sector, 4 conclusions, 5 future scope, authors’ contributions, conflict of interest statement, data availability.

• < Previous

Applications of artificial intelligence in power system operation, control and planning: a review

• Article contents
• Figures & tables
• Supplementary Data

Utkarsh Pandey, Anshumaan Pathak, Adesh Kumar, Surajit Mondal, Applications of artificial intelligence in power system operation, control and planning: a review, Clean Energy , Volume 7, Issue 6, December 2023, Pages 1199–1218, https://doi.org/10.1093/ce/zkad061

• Permissions Icon Permissions

As different artificial intelligence (AI) techniques continue to evolve, power systems are undergoing significant technological changes with the primary goal of reducing computational time, decreasing utility and consumer costs and ensuring the reliable operation of an electrical power system. AI techniques compute large amounts of data at a faster speed than numerical optimization methods with higher processing speeds. With these features, AI techniques can further automate and increase the performance of power systems. This paper presents a comprehensive overview of diverse AI techniques that can be applied in power system operation, control and planning, aiming to facilitate their various applications. We explained how AI can be used to resolve system frequency changes, maintain the voltage profile to minimize transmission losses, reduce the fault rate and minimize reactive current in distributed systems to increase the power factor and improve the voltage profile.

The demand for advanced research and technology has steadily increased in the sector of electric grids [ 1 ]. Automation and intelligent technology have become widely used in response to development demands over time. Traditional research methods are quickly becoming insufficient to enable data scientists and researchers to keep up with any global challenges that artificial intelligence (AI) may be able to assist us in solving and uncovering important insights among the billions of pieces of data scattered throughout power systems. AI can handle large amounts of data and utilize them to make power system operations, control and planning more efficient. The use of AI technology in power systems has been investigated and debated in related areas and has resulted in more study material and certain outcomes, which are reviewed in this paper. The demand for advanced research and technology has constantly risen in the sector of electricity grids. The application of AI technology to the automation of power system control can improve the efficiency of electrical automation management, mitigate the risk of accidents and ensure smooth operation of the power system over an extended period [ 2 ]. Evaluating the use of AI technology in power systems requires a comprehensive analysis of existing research in the field of artificial intelligence and its related industries.

The power system is a network consisting of three components: generation, distribution and transmission. In the power system, energy sources (such as coal, sunlight, wind, nuclear reactions and diesel) are transformed into electrical energy [ 3 ]. There are different power systems, such as solar power systems, wind power systems, thermal power plants [ 4 ], nuclear power plants, geothermal power plants, etc. All power systems have different structures and equipment for the generation of electricity [ 5 ]. The basic structure of a power system includes:

(i) generating substation;

(ii) transmission substation;

(iii) sub-transmission substation;

(iv) distribution substation.

In power system problem-solving, conventional approaches such as practical numerical optimization methods (e.g. lambda iteration and Newton–Rapson methods) have been used. Optimization problems are non-linear and, with the various constraints included, these optimization problems become slow and complex. So, several AI techniques are discussed here to solve many optimization issues with less computation time. Furthermore, experiments were conducted to identify which backpropagation algorithm would give the most efficient and reliable network training [ 6 ]. The systematic approach was introduced [ 7 ] for developing a recurrent neural network (RNN) that could effectively predict the dynamic behaviour of a pilot-scale entrained-flow gasifier. The RNN was trained using a data set consisting of input and output data collected from a dynamic read-only memory (ROM) system established in a previous study. To evaluate the accuracy of the RNN, comparisons were made with computational fluid dynamics models and experimental data obtained from the pilot-scale gasifier. The findings demonstrated that the RNN surpassed the performance of the dynamic ROM model, showing strong predictive capabilities to capture the transient behaviour of the gasifier. Additionally, this was achieved while significantly reducing the computational resources required. Power systems are one of the main study topics for the advanced development of AI. The use of AI algorithms in power plants has been under continual investigation since the advent of the expert system technique [ 8 ]. However, in most situations, issues such as long cycle times, complex computation and difficulty in learning arise with classic AI methods. In recent years, efficiency has increased significantly with the continued advancement of AI algorithms. The multi-source model of heterogeneous large data has gradually developed, the data volume continues to grow, and new possibilities and problems are being created with the use of AI in power systems. AI encompasses several technologies such as expert systems, pattern recognition, genetic algorithms (GAs) and neural networks. By incorporating AI into the automation of power system control, it has the potential to enhance the efficiency of electrical automation management, mitigate the risk of accidents and ensure long-term smooth operation of the power system. Machine learning (ML) has also found extensive applications in predicting the properties of rechargeable battery materials, particularly electrolyte and electrode materials, as well as the development of novel materials, according to relevant research. The breadth of ML applications will grow steadily as ML technology advances and new unique issues emerge in the research of rechargeable battery material [ 9 ]. Although ML has shown considerable potential in modelling complex systems, its implementation introduces new challenges. These challenges include difficulties associated with accessing relevant and reliable data sets, and addressing inaccuracies in model predictions needs to be addressed before ML can be widely deployed. To effectively implement ML on a large scale, it is crucial to have compelling evidence of its effectiveness in diverse areas such as manufacturing processes, energy generation, storage and management. Furthermore, the availability of commercial software and a skilled workforce specialized in the relevant domains is essential [ 10 ]. ML has proven to be beneficial in creating data-driven models that accurately correlate material properties with catalytic performance, including activity, selectivity and stability. As a result, there have been advancements in the development of effective design and screening criteria for solid-state catalysts with desired properties [ 11 ]. Nevertheless, there are still difficulties in applying existing ML algorithms to accurately predict catalyst performance or devise strategies for designing high-performance catalysts. The review emphasizes recent advancements in ML applied to solid heterogeneous catalysis, as well as the limitations and constraints faced by ML in this field, and also discusses some of the prospects for using ML effectively in the design of solid heterogeneous catalysts. Successful uses of ML in short-term hydrothermal scheduling will strengthen the link between real operations and issue formulation, and prepare the hydropower sector for autonomy by identifying the need for and availability of autonomous systems now and in the future. In this research, a review of the state of the art of ML applications for the hydroelectric sector was offered [ 12 ]. Digital technologies have a significant impact on energy market services and the safety of residents and energy consumers, particularly in smart homes. Sustainable smart home networks can improve energy efficiency, utilize local renewable energy, decarbonize heating and cooling systems, and promote responsible electric vehicle charging [ 13 ]. The next decade is crucial for achieving ambitious global CO 2 reduction targets and the decarbonization of buildings is a major challenge. Water-efficient development and resilient homes are essential for coping with impacts of climate change. Research on sustainability and energy efficiency is vital to improve the quality of life in the face of climate change [ 14 ]. Several important aspects were highlighted [ 15 ] with respect to the current state and prospects of smart homes. It was recognized that despite the increasing prevalence of smart homes and the growing familiarity with them, there were still significant obstacles that researchers must address to achieve widespread adoption [ 16 ].

One of the technical hurdles highlighted is the diversity of manufacturers and devices, each with varying charging systems, frequencies and communication methods [ 17 ]. This fragmentation can hinder interoperability and compatibility between devices and systems. The greater acceptance of smart home technologies poses a significant challenge. This review highlights the crucial task of convincing consumers about the safety and reliability of these technologies. An approach [ 18 ] was presented to develop prediction models that were capable of identifying faults and malfunctions in power equipment, demonstrating their effectiveness in predicting the progression of degradation phenomena. The challenges were discussed [ 19 ] associated with predicting the technical condition index of the equipment and determining the probability of its current state having defects. This research contributes significantly to the advancement of predictive analytics tools in the industry, enabling proactive maintenance of equipment. ML and data-driven approaches exhibit significant promise in the field of predictive analysis within power systems, especially in the context of smart grids. These methods can efficiently analyse the vast amounts of data collected from smart meters and other devices in real time, facilitating optimized energy flow in an increasingly renewable-energy-focused landscape [ 20 ]. They offer advantages such as improved accuracy, cost reduction and improved efficiency. However, certain challenges must be overcome, such as ensuring the availability of high-quality data and managing the potential risk of information overload [ 21 ].

The articles selected for review are based on different parameters and selection criteria. The shortlist is based on parameters such as duration, analysis, comparison and applications, as listed in Table 1 . The challenges investigated for power system operations, control and planning in the article are as follows and a diagram visualizing the domains of the power sector along with the AI techniques used and their application is presented as Fig. 1 . Power system operation [ 22 ] includes the total power requirement that must reliably meet the real-time generation, including transmission losses. The problems involved in this task are economic load dispatch (ELD), power flow, unit commitment and generator maintenance schedule.

Selection criteria for shortlisted research papers

Visual depiction of power sector domains, their application and AI techniques used

The complex and large design of the power system is presented [ 23 ] and interferences in the power system are a problem. When a large interference occurs, control tasks are needed to find the disturbed area, control the impact caused and bring the process to normal form. Heuristic solutions are non-linear and hence are not designed to deal with fast-occurring disturbances. Therefore, many control optimization techniques such as voltage control (VC), power system stability control and load frequency control are discussed to address this problem.

Power system planning has an arrangement of a power system that is complex and large with many parts such as flexible alternating current transmission system (FACTS) devices and distribution systems. The major goal of least-cost planning is to optimize the components required to deliver enough power at a minimal cost. Many factors such as FACTS placement and demand are given importance in the expansion of power system planning. Reactive power optimization, distribution system planning (DSP) [ 24 ] and capacitor placement are the optimization problems considered in this task [ 25 ].

2.1 Artificial neural network

In AI, a set of inputs is transformed into an output using a network of neurons. A neuron produces a single power by simply operating its input in the same way as a processor [ 26 ]. The working group of neurons and the pattern of their connections may be utilized to build computers with real-world issues in model recognition and pattern categorization. As the human brain processes are replicated, input signals are processed using mathematical operations utilizing artificial neurons.

The network consists of neurons organized in layers and connected to ensure information input–output flows [ 27 ]. By using what is known as the activation function, in layer ‘i’, each neuron is linked to the ‘i+1’ layer of all neurons. The input signals for a specific neuron originate from all neurons in the prior layer and their excitation power changes to govern the degree of signal reaching each neuron [ 28 ].

In several scientific disciplines, such as medical diagnosis, voice, pattern recognition, etc., artificial neural networks (ANNs) are utilized. The ANN is a computing system partly based on biological neural networks, expressed by linked nodes (artificial neurons), correctly structured in layers that are found in human or animal brains. All artificial neurons are linked and are able, employing their connections (synapses) to send signals, generally real values, which result in an output computed according to the original input, depending on the sizes allocated to all neurons [ 29 ].

ANNs are recognized as data-mining approaches capable of modelling several independent characteristics with dependent functions in non-linear functions. ANNs may predict a future value of a dependent variable after training with a comparable sample, replicating the learning process of a human brain [ 30 ].

In turn, a difference in signal strength affects the activation of the neuron and, as a result, signals that are transmitted to other neurons as shown in Fig. 2 .

Simplified diagram of the artificial neural network

Input layer—distribute other units but does not process the data.

Hidden layer—the ability to map the non-linear problems is provided through hidden layers.

Output later—the output units encode the value to be assigned to this instance.

2.2 Adaptive neuro-fuzzy interference system

The adaptive neuro-fuzzy interference system (ANFIS) creates an input/output data set whose membership function parameters are modified with the minimum square method type or the backpropagation algorithm by itself, using a fuzzy interference system [ 31 ]. This modification has helped the fuzzy system learn from the data it models. By applying hybrid learning, ANFIS utilizes a systematic approach to determine the optimal distribution of membership functions, enabling effective mapping of the relationship between input and output data [ 32 ]. The ANFIS architecture combines ANNs with fuzzy logic, making the modelling process more structured and less dependent on expert knowledge. This inference system is constructed using five layers in its basic form. Each ANFIS layer has several nodes defined in the layer specification using the node function. The current layer inputs from the preceding layer nodes are collected. The structure of the ANFIS is shown in Fig. 3 [ 33 ].

Simplified diagram of an artificial neuro-fuzzy interference system

The fuzzy-inference method involves organizing empirical information in a professional manner, which presents challenges in quantifying it through membership functions (MFs) and fuzzy rule bases [ 34 ]. Additionally, neural networks possess learning capabilities. From top to bottom, they are very adaptable in their system set-up and have great parallel processing and fault tolerance. The theories for neural network neuro-fuzzy systems are actively explored in several areas [ 35 ].

The utilization of a neuro-fuzzy system, which emulates human learning and decision-making abilities, can lead to varying model performances compared with traditional mathematical approaches. The process of rule generation and grouping in a neuro-fuzzy inference system, adapted to the specific model, can be approached through a grid-based methodology, known as the ANFIS.

2.3 Fuzzy logic

To identify the fuzzy set from which the value comes and the degree of membership within that set, fuzzy logic systems base their choices on input in terms of variables generated from the member functions [ 36 ]. The variables are then combined with IF–THEN language requirements (fuzzy logic rules) and a fluid implication is used to answer each rule [ 37 ]. The response of each rule is weighted according to the confidence or degree of the inputs of each rule and the central part of answers is computed to provide a suitable output and achieve the compositional rule of deference. Now, the design of fuzzy logic systems is not a systematic approach. The easiest way is to subjectively define member functions and rules with a human-operated system or an existing controller and then test the design for the right output. If the design fails the testing, the MFs and/or rules should be changed. Recent investigation directions involve the creation of fuzzy logic systems that can learn from experience.

Currently, only published findings can create and modify fuzzy control rules based on experience [ 38 ]. Among them can be Scharf’s self-organizing robotic control system [ 39 ] by using a performance matrix to change the rule matrix and alter the rules that constitute the management strategy. Another intriguing example is the Sugeno fuzzy vehicle, which can be trained to turn and park itself. Instead of a membership function, the effect of a rule is viewed as a linear equation of the process state variables. Through optimization of least-squares performance indices using a weighted linear regression system, the challenge is simplified to a parameter estimate. Although these approaches provide promising outcomes, they are subjective, somewhat heuristic and depend on trial and error for the choice of MFs. Thus, the ability to learn neural networks can offer a more promising approach to fuzzy logic systems [ 40 ]. As shown in Fig. 4 , the fuzzy logic system consists of four parts: knowledge base, fuzzification, inference and defuzzification. On the basis of the fuzzy constants provided, the process converts given inputs to the fuzzification stage. Based on the knowledge base, the inference is made. Then, in the defuzzification stage, every fuzzy output is mapped to complex output MFs [ 41 , 42 ].

Fuzzy logic system

2.4 Ant colony optimization

Ants exhibit a behaviour in which they remember and follow a specific path between their colony and a food source. They achieve this by leaving pheromone trails during their food search [ 43 ]. When other ants come across these pheromone trails, they start to follow them. Because the increased presence of the chemical on the path has the effect of attracting more ants to follow it, the ants will emphasize the pheromone trail. To find the best solution to the problem studied, ant colony optimization (ACO) builds multiple iterative solutions. The objectives of [ 44 ] were to evaluate the features of the search area for problems and to use this knowledge to address the solution process. The solution–construction process is a sequential decision-making process due to parametric stochastic decisions. An ACO algorithm depends on a sequence of learning of the parameters used in decision-making to reach a global policy that provides optimum solutions for a particular situation [ 45 ]. The parameters of the learning object are considered pheromones and are called variables of the pheromones.

An ACO method contains a stochastic local search technique to organize the routing pathways that artificial ants can determine. These ants co-operate together through indirect information exchange to create the best and shortest route. The concept of the ACO is taken from the food search characteristic of the true colony in an intelligent optimization algorithm and how the ants work together in this difficult job. It can be expected that the ACO finds the quickest route from nest to food according to the biological study of the ants. The ant pheromone distribution technique is termed staggered, in which information is shared with other ants indirectly. Pheromone updates are the basis of the ACO algorithm. These pheromone updates depend on the pheromone and the number of ants that work best. Natural ants can determine the quickest route based on their best knowledge and a strong pheromone trace. The shortest path is inversely proportional to the amount of pheromone and length of the path using an ACO method. The following is a step-by-step explanation of the algorithm replicating these properties [ 46 ]. The pseudocode for ACO is shown in Table 2 .

Pseudocode of the ant colony optimization algorithm

Set pheromone pathways: The algorithm starts by setting the initial pheromone pathways in the search space of the problem. These pathways act as a guide for ants to navigate and find solutions.

Generate a random ant population: Next, the algorithm generates a population of random solutions (ants) to start searching for the optimal solution.

Choose the optimal position: Each ant then uses a combination of pheromone information and heuristics to determine the next step (position) to take. The objective is to find the position that maximizes the target function.

Get the finest search ant: After all ants have completed their search, the algorithm selects the best ant, i.e. the one with the highest value of the target function.

Restore the trail of pheromone: The pheromone trail of the best ant is then updated to reinforce its path, encouraging other ants to follow it.

Check the end condition: The algorithm repeats the above steps until a stopping criterion is met.

End: The algorithm concludes when it satisfies the stopping condition and provides the best solution discovered.

2.5 Artificial bee colony optimization

The artificial bee colony (ABC) optimization imitates bee behaviour. A colony of bees is made up of onlookers, scouts and worker bees [ 47 ]. Artificial bees are flown in this system in a multidimensional search room and, depending on the experience they have gained and based on their next partner experience, the used bees pick their food sources and bees to change positions. Scout bees pick their food sources at random without any experience. Each food source avoids the probable solution to the problem under discussion [ 48 ]. The number of bees employed is as high as the food sources, each being a site, which is currently being used or as many solutions as individuals [ 49 ]. This procedure is continued until the ABC optimization meets a stop criterion.

ABC_Optimization (n, m, k)

population <- initialize (n, m, k)

global_best <- assign_random_food_source(population[m])

while! stop_criteria_met ()

for bee in population

fitness <- calculate_fitness (bee. food_source)

if fitness > global_best. fitness

global_best <- bee. food_source

for the bee in population

bee. update_food_source (global_best, bee. next_partner)

update_food_source (global_best, next_partner)

prob <- random_probability ()

if prob < experience

food_source <- global_best

else if prob < experience + next_partner. experience

food_source <- next_partner. food_source

food_source <- random_food_source ()

Initialization phase:

Initialize the x i j solution population in the j domain parameter. The exact description may be used for that purpose:

where x m a x j is the upper bound of the parameter j and x m i n j is the lower bound of the parameter j.

Worker bee phase:

Each worker bee uses a formula to identify and assess a food source v i j representative of a location such as a food source in her x i j memory. Each worker offers information about their food source to onlookers who select a food source website based on information collected from their bees while they wait at the hive according to Equation (2) :

If x k is a randomly picked solution, j is a parameter randomly selected and β i j is a random integer within the [–a, a] range. A greedy selection between v i and x i is applied after the production of a new solution v i ⁠ .

Onlooker bee phase:

There is a reference previously to the proportion of the amount of a food source to its location in the solution. Onlookers are positioned at food sources using a selection strategy based on fitness, such as the way of selecting the roulettes wheel. New solutions x i based on pi are picked to assess the new solutions v i and new solutions v i for spectators are created. The hired bees between v i and x i receive a greedy selection.

Scout bee phase:

Former workers who lost their resources start scouting randomly for food supplies. Every colony of bees has scout bees. The scouts have little instruction when looking for food. They mostly focus on finding food. Artificial bees can find the available answers rapidly. ABC decides that the artificial scout is the bee whose food supply has been lost or whose profitability has fallen below a specific level of profitability. The control parameter that decides the class is the withdrawal criterion or the ‘limit’. After a predefined number of attempts, a worker bee leaves an unimproved solution that is a source of food. The number of tests necessary to release the answer is determined by ‘limits’.

2.6 Particle swarm optimization

Particle swarm optimization (PSO) is a population-based evolutionary computational technique that is employed to address stochastic troubleshooting. It belongs to the category of swarm intelligence and is founded on social and psychological principles. PSO provides valuable insights into engineering applications and contributes to their development [ 50 ]. Social impact and social learning make cognitive consistency possible for the person. People may resolve issues by talking to people and by changing their ideas, attitudes and behaviour; they can usually be portrayed as people moving in a socio-cognitive space towards one another. But PSO has certain inconveniences such as global convergence; unlike some other optimization algorithms, PSO does not have a guarantee of global convergence, which means that it may not find the true optimal solution. To address this drawback, a novel PSO and a chaotic PSO are designed to tackle energy-system optimization issues efficiently. The analysis of the problem of unit commitment within the regulated system leads to the examination of UCP (uniform customs and practice for documentary credits) inside the deregulated market. The overall profit, execution time and convergence criteria are compared between various approaches.

One element is the current velocity of the particle v ( t ) ⁠ . Another is the optimum position Y ∗ ( t ) to approach the particle. The third factor is that the community or sub-community is optimally informed by Y ∗ ∗ ( t ) [ 51 ]. In each iteration step, the particle speed is modified to Y ∗ ( t ) and Y ∗ ∗ ( t ) ⁠ . Meanwhile, the random weight is independently allocated to the V i ⁠ , Y ∗ ( t ) and Y ∗ ∗ ( t ) ⁠ . The speed and position are updated following Equations (3) and (4):

In the given equation, v k +1( i , j ) represents the velocity of the particle in the i -th particle and j -th dimension at iteration k  + 1.

The weight factor ω determines the extent to which the previous velocity influences the new velocity.

v k ( i , j ) denotes the velocity of the particle in the i -th particle and the j -th dimension at iteration k .

C 1 and C 2 are the learning parameters that determine the influence of the personal best and global best solutions, respectively.

r and 1 and r and 2 are randomly generated numbers within the range of [0,1].

P bes t k ( i , j ) represents the personal best position of the i -th particle in the j -th dimension achieved thus far.

Y k ( I , j ) represents the current position of the i -th particle in the j -th dimension.

G b es t k signifies the global best position discovered by all particles up to the present iteration.

The flow chart for PSO is shown in Fig. 5 .

Flow chart of particle swarm optimization

2.7 Regression model

The research model [ 52 , 53 ] can be defined using Equation (5) :

where Y represents the dependent variable; this refers to the indication of respondent i ’s willingness to adopt smart home technology and their level of flexibility in terms of demand for technology j. β refers to the intercept. X 1 ij ,..., Xnij are dichotomous predictors included in the model. εij represents the random error term.

Building on Equation (5) , the level 2 model can be formulated as follows:

In Equations (5) and (6) , u 0 j ,..., u 1 j represent the random effects. W 1 j and W 2 j correspond to grand-mean centred and uncentred variables, respectively.

These equations are utilized in research to describe the relationships between the dependent variable, predictors and random effects. Equation (5) serves as the core model equation, capturing the influence of the predictors on the dependent variable while accounting for random error. Equations (5) and (6) extend the model by specifying the relationships and random effects associated with the intercept and predictor coefficients at the level 2 analysis. Collectively, these equations offer a comprehensive framework to analyse the variables that impact the acceptance of smart home technology and the adaptability of demand within the specific research context.

2.8 Regression and classification problems using AI

The RNN is a variation of the neural network frequently employed in the power systems domain to address regression and classification problems that involve sequential data. Unlike direct neural network models, the structure and operating principle of the RNN differ significantly [ 54 ]. In an RNN, the input data are fed to the model sequentially at each time step ( t ), as shown in the signal propagation diagram. At each step, the current state ( output ) is calculated by considering the current input data and the previously computed state. This iterative process continues for a fixed number of steps ( n ) until the desired output (predicted value) is achieved or until all input data ( input ) have been processed [ 55 ].

The propagation of signals in the RNN model is illustrated by the values assigned to each hidden state (hidden). These hidden states are calculated using the previous hidden state ( hidden  − 1) and current input data ( input ) [ 56 ]. hiddent = (〈 w hidden , hiddent – 1〉 + 〈 w input , input 〉) Here, σ () represents the activation function (such as the sigmoid function, hyperbolic tangent or rectified linear unit (ReLU)), while w hidden and w input are the weights for the hidden and input states, respectively.

The output value at each calculation step output is obtained by taking the dot product of the weights associated with the output state and the values of the hidden state, similar to a regression equation: output = 〈w output , hiddent 〉 [ 57 ].

During training, the initial stage involves calculating the output signal, after which the error function is calculated to determine the discrepancy. For regression problems, it is common to utilize the square root of the standard deviation between the output of the RNN and the values from the response space ( y t ):

Applying the chain rule, the gradient of the error functional is calculated. The weight coefficients ( w ij ) are adjusted in a manner that reduces the functional, following the direction of decreasing values, until it reaches the minimum value or the training iterations reach the predetermined limit. It is important to note that the weights associated with the hidden state of the RNN ( w hidden ) remain unchanged after propagating the error from each output ( output ). Conversely, the coefficients w output and w input change at each step of the gradient [ 58 ].

3.1 Operation of the power system

ELD is the process of assigning the generation output to the generation unit to supply the system load fully and economically. The whole engaged generating unit produces total electricity costs to minimize energy. The main concern of the ELD problem is to reduce the overall fuel cost. This is achieved by generating electricity in a way that optimizes the use of resources and reduces the overall cost of power production in the electricity system. Multiple generators provide enough total output to meet the consumer requirements in a typical power system. The production costs of each generating unit in the electricity system are different, as the producing units are not the same distance from the loading unit. Over the years, several AI approaches have been created to address this challenge. After simulation, authors concluded that using the genetic algorithm technique to solve the ELD problem can result in a lower overall cost of electricity production, but may also result in higher emissions. However, choosing a solution with a higher cost may result in minimum emissions. Transmission losses are usually neglected when they are small. However, for long-distance transmission in large and interconnected networks, transmission losses become significant and have an impact on the optimal distribution of power generation. It is possible to operate the same multitasking system with a better voltage profile and with evolutionary calculation technology, the cost and emission value of the best compromise.

The proposed results of the simulation of the ANN emphasize that the results are indiscernible from conventional methods, although the time used by neural networks is less than conventional methods. The number of generators increases the prediction error because there is a lot of input and output data to be learned. To monitor the performance, neural networks have been modelled. The authors have performed the simulations with many generation units having ramp rate limits and prohibited operating zones as constraints and the resultant performance is compared with ANN, GA and ACO techniques, but the ABC technique gives better outputs with fast convergence. The greedy selection procedure and the timely abandonment of the used food sources contained in ABC give it this potential. The basic operations of ABC optimization prevent solutions from stopping and make the algorithm more exploitative.

3.1.2 Generator maintenance scheduling

Generator maintenance scheduling (GMS) is a complicated combinatorial optimization issue for a power provider. Mathematical approaches include traditional ways to tackle the GMS issue. To evaluate the needed objective function, a mathematical model approach employs a trial-and-error procedure. Mathematical approaches even fail to provide viable answers as in some circumstances the operator needs to rely on certain assumptions and models that may not accurately reflect real-world conditions. In some cases, operators may need to be involved to provide additional input and expertise to make informed decisions about maintenance schedules. In addition, there may be unpredictable factors, such as equipment failures or changes in demand, that cannot be accurately accounted for by mathematical models alone. Maintenance is a preventive outage program for generating units within a certain time horizon in a power system. In the event of a range of various specification generating units in the energy system and several limitations to produce a sustainable and practical solution, maintenance planning becomes a difficult challenge. The maintenance planning of the generators is done for time horizons of different lengths. Short-term maintenance plans from 1 hour to 1 day are crucial to the daily operations, engagement and operational planning of power plants. Medium-term planning is necessary for resource management between 1 day and 1 year. Long-term planning from 1 to 2 years is crucial for future planning. An examination is being conducted to resolve some AI methods, including simulated developments, neural networks and GAs. The application of the genetic algorithm through case research shows that suitable GA parameters are safeguarded, as well as issue coding and development functions. The use of integer encoding decreases the velocity of the genetic search method investigation. By using integer encoding, the algorithm needs to perform additional operations to convert solutions into integers, which can slow down the search process. Planning the generation of power remains a barrier to competent solution technology and a difficult optimization problem. The challenge in power generation planning lies in finding the optimal balance between cost and efficiency, while also considering factors such as environmental impact, reliability and security of supply. The answer to the difficulties in generation planning consists of finding the UC (unit commitment) at every point in the programming period for each generator in one power system. An electrical system must be defined in each planning interval for each power generator for the decisions and levels of output. The solution process must be addressed concurrently for binary decisions and continuous variables. Generation difficulties with scheduling are typically quite narrow and combined. Match swarm optimization approaches have been used to achieve viable schedules within a specified time. The study found that an optimization-based approach using PSO provided better results than a GA or an evolutionary strategy. Data from the actual power system were used to evaluate the performance of the different optimization techniques.

UC is properly scheduled for the ON/OFF status and the genuine generator power outputs of the system. To satisfy a high number of system limitations and decrease the overall fuel cost at every time interval, a spinning reserve is necessary (spinning reserve refers to the additional generation capacity that is available and running but not actively supplying power to the grid). UC meets the expected load requirements in advance. To implement UCs, medium-term load forecasting using ANN was used. The neural network structure was trained through learning and parameter learning. Total operational expenses under 24 hours were used for the assessment criterion. The study demonstrates the effectiveness of the proposed approach by comparing the performance of the ANN-based load forecasting model with traditional methods such as linear regression and time-series forecasting. The results show that the ANN-based load forecasting model significantly improves the accuracy of load forecasting and reduces the scheduling cost by reducing the number of units needed for scheduling. The study also highlights the importance of considering the uncertainty and variability of load demand in UC scheduling and suggests that ANN-based load forecasting models can be a useful tool for achieving more efficient and reliable scheduling in power systems. Locational marginal prices have been evaluated through a trained ANN. The findings show that the current technique gives a different UC mechanism. To develop unit commitment, the PSO technique is used. On implementation, with the increasing size, the execution time is also increasing. To accelerate the PSO, a convergence repair method is also implemented.

3.1.4 Optimal power flow

Optimal power flow (OPF) is a highly important technique to identify the optimum control parameter settings that enhance or decrease the intended target function, but also under a variety of limitations. An essential instrument to design and operate a power system is the issue of optimum power flow to identify the best parameter settings that can maximize or minimize the intended goal function within specific limitations. Voltage and reactive controls, also called OPD, are an OPF sub-problem that seeks to reduce overall transmission loss by resuming the reactive power flow. Optimal reactive power dispatch is a non-linear solution for the issue of blending integers since some control variables such as tap ratios for transformers, shunt capacitor outputs and reactors are distinct.

The alternate strategy for mitigating the problem of GA-ANN is set out in this article. A collection of ANN networks is trained offline in specified system quantities to work on a general OPF issue. To choose the appropriate ANN inputs, the k-mean clustering technique is utilized. When learning the functions correctly, ANNs can easily estimate the associated results with great precision.

The ANFIS develops the input/output data set fuzzy-inference system (FIS) that matches the membership (adjustment) parameters with a backpropagation or minimum square process type. This update allows you to learn from the fuzzy systems data IEEE 39 bus system implementations and simulated software from the power world are utilized for the formation of the ANFIS. The results indicate that the ANFIS offers solutions as accurate as conventional ones. It takes less time, though, and it works quickly. Some additional papers on the application of AI in the operation of power systems are presented in Table 3 .

Applications of artificial intelligence in the operation of a power system

3.2 Control of the power system

The main objective of a voltage controller power system is to maintain the voltage profile within a defined limit, thus minimizing transmission losses and avoiding cases of voltage instability [ 81 ]. The VC system consists of three levels of hierarchical control: AVR (automatic voltage regulator), tertiary voltage control (TerVC) and secondary voltage control (SecVC). AVR is aimed at controlling the voltage of buses that are equipped with reactive power sources (e.g. synchronous, sync, static var compensators and STATCOM (static synchronous compensator)). Actions are carried out locally at this control level. SecVC is used to monitor the voltage on a specific bus that controls a cargo bus.

In situations in which there is hardware present in the vicinity that modifies the reference point for the AVR, the control level typically operates at a slower pace compared with the AVR control level. SecVC is responsible for identifying VC regions and their correlation with individual load buses. To accommodate varying power system conditions, SecVC must demonstrate flexibility in adjusting the control regions to accommodate all grid conditions. On the other hand, TerVC determines the optimal reference value for voltage grids at each load bus. The objective is to minimize power loss, optimize reactive power and maintain a minimum load release or reservation. TerVC is usually updated every 30 minutes to 1 hour.

The backward error propagation algorithm trains the multilayered feedforward perception. The minimum singular value method analyses the static voltage collapse. The procedure uses a minimum voltage stability evaluation time once the network training is complete. For monitoring voltage collapse, complementary methodologies of neural networks and expert systems would be combined for use in the application [ 82 ].

GA is an iterative optimization technique with several solutions from the candidates (known as a population). In the case in which there is no knowledge of the problem field, then the GA starts to look for solutions from a random population. The appropriate coding (or display) must first be defined to solve the problem. A fitness function should also be defined so that every coding solution is given a figure of merit. If parents are not satisfied with the termination condition, for reproduction, they must be picked [ 83 ]. They are then joined to generate offspring through reproduction and, to refresh the population of candidate solutions, crossover and mutation operators are utilized. Typically, in a basic genetic algorithm, three operators are involved: selection, crossover and mutation. These operations are performed to generate new offspring, individuals and subsequent generations. The same process is repeated with the new generation until the desired criteria are met. The approach of this method is used to teach swarming at the beginning of PSO. In this case, 10 control variables are used for the ANN input. The neuron and its prejudice are 11. A hidden layer consists of this group of neurons and biases. Ten outputs/goals are available. These objectives are achieved by using the optimal value of PSO. The last outputs are the initializations in the time-varying non-linear particle swarm optimization (TVNL-PSO). The steps are as follows:

PSO is used to take the ANN input; the weight value is applied at random;

the ANN input and partition in a cached layer are weighed and then activated by the sigmoid binary function;

weighting of the output in the hidden layer and activation of the linear function;

to optimize the reactive power and VC by TVNL-PSO, the ANN output is transmitted as a starting initialization value.

3.2.2 Power system stability control

The stability of a power system is a feature that allows it to remain under a balance in normal operating conditions and retrieve an acceptable balance after a change. Margins of stability can be seen to decline throughout the world [ 84 ]. We highlight three of the many reasons for this:

The inhibition by economic and environmental constraints of further transmission or construction. Therefore, power systems must be operated with lower safety margins.

Restructuring of the electricity industry. The restructuring process reduces the margins of stability, as power systems do not co-operate effectively [ 85 ].

Increased complexity of power systems multiplies the compulsive properties. These include large, non-linear oscillations; frequency differences between weakly binding energy-system areas; interactions with saturated devices.

Fuzzy logic endeavours to address problems by emulating human reasoning, allowing optimal decision-making based on available information. It can also be employed to regulate the stability of un-modelled systems. To achieve improved performance, a fuzzy logic (FL) controller is combined with a PID (proportional–integral–derivative) controller. In this particular scenario, the fuzzy logic control adjusts the gains of the PID controller based on the power system.

A fuzzy logic controller primarily consists of four major parts: fuzzification, fuzzy rule base, fuzzy inference and defuzzification. FACTS have proven to be extremely promising for increasing performance under stable conditions. The most promising FACTS device is a unified power flow controller (UPFC). Three control factors can be adjusted: bus voltage, reaction line and phase angle between two buses. The power flow should be redistributed across lines while a stable state is preserved. It can also be utilized to increase damping when low frequencies are damped temporarily.

Power system stability control.

Load frequency control as defined by the controllable generator power output control in a prescribed area resolves system frequency changes, two-line loadings or interactions to maintain an interchange with other regions within the fixed limit or scheduled system frequency [ 86 ]. The traditional proportional–integral (PI) controller is the most widely used among different types of load frequency controllers. The PI controller can be easily implemented and provides a faster response, but its performance decreases when unwanted disturbances, such as load change dynamics, increase the difficulty within the system. In this paper, less computing is required for the non-linear autoregressive-moving average-L2 (NARMA-L2) control architecture. Plant output, reference and control signals are included. The controller is therefore taught to monitor the output of the reference model. The model network that updates controller settings predicts the effect of the change in plant performance. Some additional papers on the application of AI in the control of power systems are presented in Table 4 .

Applications of AI in control of a power system

3.3 Planning of power system

DSP plays a crucial role in enhancing reliability and minimizing costs for both utilities and consumers. Electric power distribution networks are a fundamental component of the electrical power system. In general, transport networks are denser and more complex than those that provide transformer stations [ 110 ]. Automating previously manual jobs increases with distribution networks becoming more complex. New tools are known as advanced automation functions that support the operation of such networks. These functions enable the network operator to effectively address issues that arise. Furthermore, the reconfiguration of distribution networks is essential to identifying optimal solutions that align with the operator’s requirements and constraints, ensuring a secure and economically optimized electricity supply.

The optimal design of a distribution network is not a fixed solution, but rather a process that involves considering various technically feasible options and using improvement tools to make the best decision based on factors such as demand, reliability of power transmission and network structure. All potential paths are initially identified with uploaded system data and then the energy-loss cost calculation applies for each identified path forward/backward sweeping load flow technique. For the distribution of power, the minimum energy-loss path is chosen. The optimal selection of the branch conductor of the radial system is done using optimization of PSO. In this case, parameters such as power loss, voltage profile and capital investment depreciation improve optimization. These parameters are used as optimization criteria to determine the optimal branch conductor that minimizes power loss, improves the voltage profile and reduces capital investment depreciation. The PSO algorithm iteratively updates the position of each particle in the search space based on its own experience and the experiences of its neighbours. The algorithm continues until a global optimum is found or a stopping criterion is met. The final solution produced by the PSO algorithm represents the optimal branch conductor that meets the optimization criteria [ 111 ]. The optimization of the PSO results in the optimal conductor and the best substation, the positioning of the optimal conductor is selected and then the optimal substation power distribution is achieved.

A multi-target algorithm was proposed that uses a fluid optimization technique to handle contradictory targets [ 112 ]. The plan formulation and the algorithm include a multi-target function that uses battery energy storage systems (BESSs) and traditional resources to select the best planning option. The microgrid BESS has been receptive to power management and improvement in power quality. The proposed algorithm is based on the fuse-based decision-making processes of the Mamdani-type FIS and Bellman–Zadeh approach.

In this paper [ 113 ], two algorithms, namely the mixed-integer linear program (MILP) and GA, are compared for the design of a radial distribution system feeder. The main objective is to minimize total investment and operational outages while maximizing system reliability. The study aims to evaluate and compare these two optimization techniques in terms of their optimality, complexity and time requirements. A unique aspect of the optimization model is the consideration of operational costs associated with failures, which are directly linked to the design of the system. The fault rate and defect cost at each loading point are updated based on the proposed configurations. It is crucial to determine which method produces superior results in terms of optimality, complexity and time efficiency.

The GA technique is used to build the algorithm for optimizing distribution networks. The fundamental concept is the growth of the genetic operator population (selection, crossover and mutation). These are used to generate a fresh population from the previous generation throughout each generation procedure. In GA, a single chromosome shows each person. This chromosome corresponds, according to the graph theory, to a radial distribution network configuration or a particular graph twist. The chromosome group is the population. Randomly, an initial group is created as a first step in the implementation of GA. Then the encoding is applied to each chromosome. In this study, only closed branches represent the network topology. A true coding strategy was used to match each gene to the edge of the coagulation tree [ 114 ].

3.3.2 Reactive power optimization

As the demand of electricity increases and new lines are built, the environment and the unforeseen power fluid in the lens are reduced, it is generated by the current scenario. Effect reactive compensation control improves voltage, reduces energy loss and improves system performance under stable and dynamic conditions in weak nodes [ 115 ]. Because the complexity of power systems is constantly increasing and the network components are constantly being loaded, abnormal operating conditions such as voltage can occur more often. Therefore, it is obvious that the power system needs adequate reactive power and VC.

In pursuit of intelligent theory development, a combination of fuzzy logic and ANNs is used to determine the control strategy for transformer taps and capacitors. However, due to the increasing complexity of control variables, rapid optimization becomes challenging [ 116 ]. To address this issue, a genetic algorithm is utilized, which tackles problems associated with incorporating regulatory time as penalty terms in the objective function and determining appropriate penalty factors that affect algorithm performance. When regulatory time is a constraint, the optimization objective focuses on minimizing the total energy loss during the dispatch period. GAs, inspired by natural selection mechanics and genetics, such as inheritance, mutation and recombination, are utilized (also referred to as crossover) to optimize the solution.

The PSO method can be used for handling FACTS devices in power systems. Various reactive power problem objectives and different solutions are addressed in the interconnected power system. Solutions and comparative analyses using the FACTS device, differential evolution (DE) and PSO algorithms are presented under various loading conditions [ 117 ]. The algorithm proposed [ 118 ] employs DE to minimize generator fuel costs on FACTS devices. Additionally, the authors discuss the hybridization of DE and PSO (DEPSO) to overcome the maximum load limit. The control of reactive flow is addressed using fuzzy sets and a fuzzy feature optimization technique is introduced to optimize reactive power. The utilization of fuzzy linear programming offers an effective approach to calculating reactive power, aiming to minimize active power loss and maximize the voltage stability margin. The paper [ 119 ] explores the combination of fuzzy and GA approaches for FACTS shunt controller placement and sizing. Lastly, the focus of [ 120 ] is on the integration of fuzzy systems with GM algorithms and the PSO algorithm to tackle the OPF problem and optimize control variables. In this paper, the authors [ 121 ] focus on fluid-based reactive and voltage controls to reduce actual loss of power.

3.3.3 Capacitor placement

There are some advantages if capacitors are placed optimally, including various factors, such as maximizing energy and reducing peak power loss through the introduction of a condenser in an electrical distribution system. In the paper, a novel adaptive modified firefly algorithm is presented to address the optimal capacitor placement problem in power systems. This optimization problem involves identifying the best positions and sizes of capacitors in a power system, in order to enhance voltage stability, minimize energy losses and improve the power factor. The proposed algorithm combines the firefly algorithm with adaptive parameter settings and introduces a unique crossover operator to enhance both convergence speed and solution quality. The authors conduct evaluations on a test system and compare the performance of the algorithm with other optimization methods. The results demonstrate the effectiveness of the proposed approach in finding optimal solutions and highlight its superiority in terms of solution quality and computational efficiency compared with other algorithms [ 122 ]. Losses occur due to reactive currents in the distribution system and are therefore minimized in the right places. Shunt capacitors are used, depending on their use. A capacitor is used to improve the voltage profile, reduce losses and increase the power factor [ 123 ].

Elbaz et al. [ 124 ] have been using ANN techniques to control both capacitor banks and voltage regulators. The ANN has many input connections and all inputs are combined to determine the output capacity. The purpose of the capacitor search algorithm was to reduce total active losses in the distribution system by utilizing the capacitor banking search to address the capacitor placement problem. The operation of the ant colony was proposed to address problems related to the installation of the capacitor. The fuzzy method uses variables such as angle, current and voltage, etc. A degree for a set and fuzzy variable is determined by MFs. This degree changes from zero to one that takes zero or one as opposed to the classical methods [ 125 ]. A fuzzy logic-based algorithm is developed to minimize line loss for the placement of condensers in a radial system. The fuzzy expert system identifies the capacitor candidate nodes by compromising the possible reduction in loss between the condenser system and the voltage level. This paper [ 126 ] presents a fluffy approach to identifying the appropriate places for capacitor placement. In the design of a seamless logic to determine the optimal placement of capacitors, two main objectives are taken into account. These objectives include (i) minimizing actual power loss and (ii) maintaining voltage within acceptable limits. Fuzzy member functions are used to model voltage and power loss indices for nodes in the transmission system [ 127 ]. The suitability of capacitor placement for each node in the distribution system is then determined. This is achieved using a set of rules within the FIS. Nodes can be fitted with capacitors of greatest appropriateness.

In [ 128 ], a method based on GAs is used to identify the optimal locations and sizes of capacitors in the distribution network. The capacitor sizes are considered discrete and known variables placed on the buses of the network. Hence, the maximum losses of the distribution system are reduced. GA technology is selected as the capacitor problem is more accurately addressed in the power grid. When the search area crosses for an optimal solution, the advantages of GA are multifunctional—that is, when a locally optimal solution is found to an engineering goal, GA adapts its search to find an optimal global solution, subject to predefined search restrictions [ 129 , 130 ]. The article shows the results of the study of the best size and location for a GA-connected system using bays in Saudi Arabia in West–East regions [ 131 , 132 ]. Two formulas are proposed for capacitor positioning: (i) cost balance condenser/loss principle and (ii) total system performance cost estimates, standard analysis and verification of annual benefits, power loss and operational tension results [ 133 , 134 ]. AI is applicable for various aspects of the power system [ 135 ]. Some additional papers on power system application of AI in planning are presented in Table 5 .

Applications of AI in the planning of a power system

Application of AI in various problems of the operation, control and planning of power systems has shown good performance over conventional methods. ABC algorithms function better than other AI techniques in the ELD problem, as shown in a comparison. The ABC algorithm has the highest-quality solutions, consistent convergence and exceptional computing efficiency. Compared with traditional mathematical models, both GA and PSO techniques are superior; the PSO technique is preferable since GA replaces humans. Unlike GA, PSO models alter humans through time and, in the following generation, all people survive. The fate of each person is continually adjusted based on the global ideal point thus far. The swarming effect in PSO allows the population of particles to move collectively in the search space, facilitating a more efficient exploration of the solution space and a faster convergence to optimum search areas compared with GA. PSO is particularly useful for optimization problems with many variables and when the solution space is complex and not well understood. Additionally, PSO requires minimal computational resources compared with GA as it does not require the evaluation and selection of multiple generations of offspring. In UC, with the increasing time, the execution time is also increasing, so the combination of many AI techniques can be progressive, as it can potentially increase the efficiency and accuracy of the optimization problem. The scalability of any heuristic optimization method is a major issue. In the load flow method, the ANFIS and ABC algorithms provide efficient and accurate solutions, and the implementation of ANNs is fast and can handle missing data effectively. In VC, rather than replacing conventional methods, the focus should be on enhancing their capabilities through the integration of AI techniques, modern control theory, fuzzy technology and ANNs, along with adaptive control and expert systems. This approach, which combines current research trends with practical experience, has great potential for practical applications. In load frequency control, the NARMA-L2 ANN network architecture provides the best output after some series of trials and improvements. Under fault conditions, it has been observed. The fuzzy-PID power output works so that the power systems are fast and stable. Due to the specialization, the fuzzy logic in condenser placement is better than other approaches. Fuzzy logic includes the relatively basic technique of calculating the necessary nodes in the distribution system to replace condensers with approximate reasoning. The results of the study indicate that the GA method can provide a globally optimal solution for reactive power optimization, particularly when an ample generation and population size is used. Furthermore, it was found that the use of the UPFC resulted in minimized losses compared with the static var compensator and thyristor-controlled series capacitor. In terms of reactive power planning, a fuzzy membership-based approach is employed in an interconnected power system to identify weak nodes and determine optimal parameter settings for FACTS devices. The feasibility of this strategy is validated through its solutions and compared with other global optimization approaches. The proposed technique is applied to a standard system under high load conditions, resulting in a stable system with reduced losses and cost savings. This approach has the potential to become a novel technology for effectively coordinating FACTS devices with other existing generators. Following the pandemic, it is expected that governments around the world will prioritize energy efficiency in buildings and smart homes. To support this, the development of open-source protocols and unified connectivity solutions is crucial. Smart home systems are focused on maximizing the energy efficiency of major household appliances, thereby contributing to overall energy balance. By constructing sustainable homes integrated with smart technologies and a combination of energy sources, significant cost savings and a reduction in carbon footprints can be achieved. While smart homes are becoming more common, there are still barriers to widespread adoption that researchers need to address. It emphasizes the potential of smart home companies and highlights technical challenges such as device compatibility. The review also discusses the importance of studying consumer attitudes and demands, and mentions the limitations of the survey methodology. Significant scientific results include an algorithm for a modern predictive analytics system, an approach to assess the condition index of the equipment and a method to determine the probability of defects using ML. The study validated the model using data from a hydroelectric power plant, demonstrating its accuracy. Future research can focus on refining the index calculation for equipment with multiple functional units and constructing predictive models for specific equipment classes. ML and data-driven techniques hold great promise in the field of power systems, especially in the context of smart grids. These methods can effectively analyse and process large volumes of data, resulting in improved accuracy and increased operational efficiency. However, some challenges need to be addressed, such as ensuring the quality of the data and interpreting the results in a meaningful way. Accurate forecasting plays a vital role in optimizing grid operations, integrating renewable energy sources and achieving cost-effective power generation. ML plays a vital role in transforming traditional grids into smart grids, improving reliability and safety. It also aids in demand-side management and enhances cybersecurity.

Evaluating energy costs and making improvements can lead to significant energy savings. Smart technologies have the potential to reduce electricity demand and environmental impact. Social acceptance of smart home systems needs to be promoted. Further research can expand sample sizes, include more diverse countries and explore smart meter readings. Future research should emphasize the importance of addressing technical, security and privacy concerns, and call for collaboration between stakeholders to enhance the smart home market. Although the developed approach offers several benefits, there is still an unresolved issue regarding the calculation of the technical condition index for equipment consisting of multiple functional units. Existing methods rely on assigning weights to each unit based on an expert evaluation to determine its importance. Additional research can focus on improving the methods for calculating the technical condition index for different types of power equipment and establishing predictive models to anticipate equipment defects in the event of functional unit failures. Future investigations should prioritize the development of more precise and dependable predictive models for power systems, taking into account the challenges related to data availability and interpretability.

Anshumaan Pathak and Utkarsh Pandey did the critical review. Surajit Mondal and Adesh Kumar supervised and reviewed the manuscript.

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

The paper is a study and review article in which no specific data are referred to. No simulation software is used.

Sozontov A , Ivanova M , Gibadullin A. Implementation of artificial intelligence in the electric power industry . E3S Web of Conferences , 2019 , 114 : 01009 .

Google Scholar

Zhao X , Zhang X. Artificial intelligence applications in power system . In: Luo X (ed). Advances in Intelligent Systems Research—2016 2nd International Conference on Artificial Intelligence and Industrial Engineering (AIIE 2016). Dordrecht, Netherlands : Atlantis Press , 2016 , 158 – 161 .

Nath RP , Balaji VN. Artificial intelligence in power systems . IOSR Journal of Computer Engineering (IOSR-JCE) , 2014 . https://jcboseust.ac.in/electrical/images/notes/aitech_ug_ai_reactive_power_control.pdf ( 29 July 2023 , date last accessed).

Das S , Mukherjee M , Mondal S. Detailed energy audit of thermal power plant equipment . World Scientific News , 2015 , 22 : 70 – 90 .

Mondal S , Mondal AK , Sharma A , et al.  . An overview of cleaning and prevention processes for enhancing efficiency of solar photovoltaic panels . Current Sci , 2018 , 115 : 1065 – 1077 .

Wasesa M , Tiara AR , Afrianto MA , et al.  . SARIMA and artificial neural network models for forecasting electricity consumption of a microgrid based educational building . In: 2020 IEEE International Conference on Industrial Engineering and Engineering Management (IEEM), Singapore, 14–17 December 2020 , 210 – 214 .

Jang JS. ANFIS: adaptive-network-based fuzzy inference system . IEEE Trans Syst Man Cybern , 1993 , 23 : 665 – 685 .

Borges AF , Laurindo FJ , Spínola MM , et al.  . The strategic use of artificial intelligence in the digital era: systematic literature review and future research directions . Int J Inf Manage , 2021 , 57 : 102225 .

Guan Y , Chaffart D , Liu G , et al.  . Machine learning in solid heterogeneous catalysis: recent developments, challenges and perspectives . Chem Eng Sci , 2022 , 248 : 117224 .

Bordin C , Skjelbred HI , Kong J , et al.  . Machine learning for hydropower scheduling: state of the art and future research directions . Procedia Comput Sci , 2020 , 176 : 1659 – 1668 .

Wang H , Ricardez-Sandoval LA. Dynamic optimization of a pilot-scale entrained-flow gasifier using artificial recurrent neural networks . Fuel , 2020 , 272 : 117731 .

Isnen M , Kurniawan S , Garcia-Palacios E. A-SEM: an adaptive smart energy management testbed for shiftable loads optimisation in the smart home . Measurement , 2020 , 152 : 107285 .

Bibri SE , Krogstie J. Environmentally data-driven smart sustainable cities: applied innovative solutions for energy efficiency, pollution reduction, and urban metabolism . Energy Informatics , 2020 , 3 : 1 – 59 .

Strielkowski W. Social Impacts of Smart Grids: The Future of Smart Grids and Energy Market Design . Amsterdam, Netherlands : Elsevier , 2019 .

Google Preview

Oliveira L , Mitchell V , May A. Smart home technology—comparing householder expectations at the point of installation with experiences 1 year later . Pers Ubiquitous Comput , 2020 , 24 : 613 – 626 .

Tao M , Zuo J , Liu Z , et al.  . Multi-layer cloud architectural model and ontology-based security service framework for IoT-based smart homes . Future Gener Comput Syst , 2018 , 78 : 1040 – 1051 .

Shcherbatov IA , Turikov GN. Determination of power engineering equipment’s defects in predictive analytic system using machine learning algorithms . J Phys Conf Ser , 2020 , 1683 : 042056 .

Moleda M , Momot A , Mrozek D. Predictive maintenance of boiler feed water pumps using SCADA data . Sensors , 2020 , 20 : 571 .

Ren C , Xu Y. A fully data-driven method based on generative adversarial networks for power system dynamic security assessment with missing data . IEEE Trans Power Syst , 2019 , 34 : 5044 – 5052 .

Yu J , Guo Y , Sun H. Testbeds for integrated transmission and distribution networks: generation methodology and benchmarks . CSEE J Power Energy Syst , 2020 , 6 : 518 – 527 .

Gonen T. Electric Power Distribution Engineering . Boca Raton, FL, USA : CRC Press , 2015 .

Wood AJ , Wollenberg BF , Sheblé GB. Power Generation, Operation, and Control . Hoboken, NJ, USA : John Wiley & Sons , 2013 .

Kumar N , Mishra VM , Kumar A. Smart grid and nuclear power plant security by integrating cryptographic hardware chip . Nuclear Engineering and Technology , 2021 , 53 : 3327 – 3334 .

Kumar N , Mishra VM , Kumar A. Smart grid security by embedding s-box advanced encryption standard . Intelligent Automation &. Soft Comput , 2022 , 34 : 623 – 638 .

Xu G , Wang Z. Power system load flow distribution research based on adaptive neuro-fuzzy inference systems . In: 2012 Spring Congress on Engineering and Technology, Xi’an, China, 27–30 May 2012 , 1 – 4 .

Karaboga D , Akay B. A comparative study of artificial bee colony algorithm . Appl Math Comput , 2009 , 214 : 108 – 132 .

Nualhong D , Chusanapiputt S , Phomvuttisarn S , et al.  . Diversity control approach to ant colony optimization for unit commitment problem . In: 2004 IEEE Region 10 Conference TENCON 2004, Chiang Mai, Thailand, 24 November 2004 , 488 – 491 .

Ho SL , Yang S , Ni G , et al.  . A particle swarm optimization-based method for multiobjective design optimizations . IEEE Trans Magn , 2005 , 41 : 1756 – 1759 .

Ratnaweera A , Halgamuge SK , Watson HC. Self-organizing hierarchical particle swarm optimizer with time-varying acceleration coefficients . IEEE Trans Evol Comput , 2004 , 8 : 240 – 255 .

Kumar N , Nangia U , Sahay KB. Economic load dispatch using improved particle swarm optimization algorithms . In: 2014 6th IEEE Power India International Conference (PIICON), Delhi, India, 5–7 December 2014 , 1 – 6 .

Panta S , Premrudeepreechacharn S. Economic dispatch for power generation using artificial neural network ICPE’07 conference in Daegu, Korea . In: 2007 7th International Conference on Power Electronics, Daegu, South Korea, 22–26 October 2007 , 558 – 562 .

Rahmat NA , Musirin I. Differential evolution ant colony optimization (DEACO) technique in solving economic load dispatch problem . In: 2012 IEEE International Power Engineering and Optimization Conference, Melaka, Malaysia, 6–7 June 2012 , 263 – 268 .

Saxena A , Pandey SN , Srivastava L. Congestion management in open access—a review . International Journal of Science, Engineering and Technology Research , 2013 , 2 : 922 – 930 .

Emami H , Sadri JA. Congestion management of transmission lines in the market environment . International Research Journal of Applied and Basic Sciences , 2012 , 3 : 2572 – 2580 .

Ghosh K , Pandey U , Pathak A , et al.  . Simulation of density based traffic control system using Proteus 7.1 professional . In: Mathur G , Bundele M , Tripathi A , Paprzycki M (eds). Proceedings of 3rd International Conference on Artificial Intelligence: Advances and Applications: ICAIAA 2022. Singapore: Springer Nature Singapore , 2023 , 493 – 504 .

Deb S , Goswami AK. Congestion management by generator rescheduling using artificial bee colony optimization technique . In: 2012 Annual IEEE India Conference (INDICON), Kochi, India, 7–9 December 2012 , 909 – 914 .

Choudekar P , Sinha SK , Siddiqui A. Transmission line efficiency improvement and congestion management under critical contingency condition by optimal placement of TCSC . In: 2016 7th India International Conference on Power Electronics (IICPE), Patiala, India, 17–19 November 2016 , 1 – 6 .

Wang X , McDonald JR. Modern Power System Planning . London, UK : McGraw-Hill , 1994 , 247 – 307 .

Dahal K , McDonald J. A review of generator maintenance scheduling using artificial intelligence techniques . In: Proceedings of the 32nd Universities Power Engineering Conference (UPEC ’97), Manchester, UK, 10–12 September 1997 , 787 – 790 .

Koay CA , Srinivasan D. Particle swarm optimization-based approach for generator maintenance scheduling . In: Proceedings of the 2003 IEEE Swarm Intelligence Symposium. SIS’03 (Cat. No. 03EX706), Indianapolis, IN, USA, 26 April 2003 , 167 – 173 .

Johnson RC , Happ HH , Wright WJ. Large scale hydro-thermal unit commitment: method and results . IEEE Trans Power Appar Syst , 1971 , PAS-90 : 1373 – 1384 .

Thum YM. Hierarchical linear models for multivariate outcomes . Journal of Educational and Behavioral Statistics , 1997 , 22 : 77 – 108 .

Reyes DM , de Souza RM , de Oliveira AL. A three-stage approach for modeling multiple time series applied to symbolic quartile data . Expert Syst Appl , 2022 , 187 : 115884 .

Almaghrebi A , Aljuheshi F , Rafaie M , et al.  . Data-driven charging demand prediction at public charging stations using supervised machine learning regression methods . Energies , 2020 , 13 : 4231 .

Huang P , Copertaro B , Zhang X , et al.  . A review of data centers as prosumers in district energy systems: renewable energy integration and waste heat reuse for district heating . Appl Energy , 2020 , 258 : 114109 .

Sun L , You F. Machine learning and data-driven techniques for the control of smart power generation systems: an uncertainty handling perspective . Engineering , 2021 , 7 : 1239 – 1247 .

Lisin E , Shuvalova D , Volkova I , et al.  . Sustainable development of regional power systems and the consumption of electric energy . Sustainability , 2018 , 10 : 1111 .

Prakash S , Jain J , Hasnat S , et al.  . Economic load dispatch with valve point loading effect using optimization techniques . In: Bansal P , Tushir M , Balas VE , Srivastava R (eds). Proceedings of International Conference on Artificial Intelligence and Applications: ICAIAA 2020, Singapore: Springer , 2021 , 407 – 416 .

Chowdhury AK , Mondal S , Mukherjee M , Biswas PK. Mega watt security assessment of power systems . International Letters of Chemistry, Physics and Astronomy, 2015 , 58 : 9 – 15 .

Dommel HW , Tinney WF. Optimal power flow solutions . IEEE Transactions on Power Apparatus and Systems, 1968 , PAS-87 : 1866 – 1876 .

Wakiru JM , Pintelon L , Muchiri P , et al.  . A comparative analysis of maintenance strategies and data application in asset performance management for both developed and developing countries . International Journal of Quality & Reliability Management , 2022 , 39 : 961 – 983 .

Muñoz-Delgado G , Contreras J , Arroyo JM , et al.  . Integrated transmission and distribution system expansion planning under uncertainty . IEEE Trans Smart Grid , 2021 , 12 : 4113 – 4125 .

Chowdhury AK , Mondal S , Alam SM , et al.  . Voltage security assessment of power system . World Scientific News , 2015 , 21 : 36 – 50 .

Holen AT , Botnen A , Stoa P , et al.  . Coupling between knowledge-based and algorithmic methods . Proc IEEE , 1992 , 80 : 745 – 757 .

Short MJ , Hui KC , Ekwue AO , et al.  . Applications of artificial neural networks for NGC voltage collapse monitoring . International Conference On Large High Voltage Electric Systems, 1994 , 2 : 38 – 205 .

Haida T , Akimoto T. Voltage optimization using genetic algorithm . In: Proc. 3rd Symposium on Expert System Applications to Power Systems, Tokyo, Japan, 1991 , 375 – 380 .

Denny FI , Dismukes DE. Power System Operations and Electricity Markets . Boca Raton, FL, USA : CRC Press , 2017 .

Masiala M , Ghribi M , Kaddouri A. An adaptive fuzzy controller gain scheduling for power system load-frequency control . In: 2004 IEEE International Conference on Industrial Technology, IEEE ICIT’04, Vol. 3 , Hammamet, Tunisia, 8–10 December 2004 , 1515 – 1520 .

Alquthami T , Butt SE , Tahir MF , et al.  . Short-term optimal scheduling of hydro-thermal power plants using artificial bee colony algorithm . Energy Rep , 2020 , 6 : 984 – 992 .

Sahay KB , Sonkar A , Kumar A. Economic load dispatch using genetic algorithm optimization technique . In: 2018 International Conference and Utility Exhibition on Green Energy for Sustainable Development (ICUE), Phuket, Thailand, 24–26 October 2018 , 1 – 5 .

Mishra SK , Mishra SK. A comparative study of solution of economic load dispatch problem in power systems in the environmental perspective . Procedia Comput Sci , 2015 , 48 : 96 – 100 .

Dixit GP , Dubey HM , Pandit M , et al.  . Artificial bee colony optimization for combined economic load and emission dispatch . In: International Conference on Sustainable Energy and Intelligent Systems (SEISCON 2011), Chennai, India, 20–22 July 2011 , 340 – 345 .

Daniel L , Chaturvedi KT , Kolhe ML. Dynamic economic load dispatch using Levenberg Marquardt algorithm . Energy Procedia , 2018 , 144 : 95 – 103 .

Ruiz-Abellón MC , Fernández-Jiménez LA , Guillamón A , et al.  . Integration of demand response and short-term forecasting for the management of prosumers’ demand and generation . Energies , 2020 , 13 : 11 .

Ali SS , Choi BJ. State-of-the-art artificial intelligence techniques for distributed smart grids: a review . Electronics , 2020 , 9 : 1030 .

Fu J , Nunez A , De Schutter B. A short-term preventive maintenance scheduling method for distribution networks with distributed generators and batteries . IEEE Trans Power Syst , 2020 , 36 : 2516 – 2531 .

Esmaili M , Shayanfar HA , Moslemi R. Locating series FACTS devices for multi-objective congestion management improving voltage and transient stability . Eur J Oper Res , 2014 , 236 : 763 – 773 .

Suresh K , Kumarappan N. Hybrid improved binary particle swarm optimization approach for generation maintenance scheduling problem . Swarm Evol Comput , 2013 , 9 : 69 – 89 .

Lakshminarayanan S , Kaur D. Optimal maintenance scheduling of generator units using discrete integer cuckoo search optimization algorithm . Swarm Evol Comput , 2018 , 42 : 89 – 98 .

Scalabrini Sampaio G , Vallim Filho ARDA , Santos da Silva L , et al.  . Prediction of motor failure time using an artificial neural network . Sensors , 2019 , 19 : 4342 .

Fikri M , Cheddadi B , Sabri O. Predicting Moroccan real network’s power flow employing the artificial neural networks . In: 2019 Third International Conference on Intelligent Computing in Data Sciences (ICDS), Marrakech, Morocco, 28–30 October 2019 , 1 – 6 .

Rahul J , Sharma Y , Birla D. A new attempt to optimize optimal power flow-based transmission losses using genetic algorithm . In: 2012 Fourth International Conference on Computational Intelligence and Communication Networks, Mathura, India, 3–5 November 2012 , 566 – 570 .

Nakawiro W , Erlich I. A combined GA-ANN strategy for solving optimal power flow with voltage security constraint . In: 2009 Asia-Pacific Power and Energy Engineering Conference, Wuhan, China, 27–31 March 2009 , 1 – 4 .

Sumpavakup C , Srikun I , Chusanapiputt S. A solution to the optimal power flow using artificial bee colony algorithm . In: 2010 International Conference on Power System Technology, Zhejiang, China, 24–28 October 2010 , 1 – 5 .

Abdellah D , Djamel L. Power flow analysis using adaptive neuro-fuzzy inference systems . In: 2015 3rd International Renewable and Sustainable Energy Conference (IRSEC), Marrakech, Morocco, 10–13 December 2015 , 1 – 5 .

Nemati M , Braun M , Tenbohlen S. Optimization of unit commitment and economic dispatch in microgrids based on genetic algorithm and mixed integer linear programming . Appl Energy , 2018 , 210 : 944 – 963 .

Alshareef A. An application of artificial intelligent optimization techniques to dynamic unit commitment for the western area of Saudi Arabia . In: 2011 Third International Conference on Computational Intelligence, Communication Systems and Networks, Bali, Indonesia, 26–28 July 2011 , 17 – 21 .

Arora I , Kaur M. Unit commitment scheduling by employing artificial neural network-based load forecasting . In: 2016 7th India International Conference on Power Electronics (IICPE), Patiala, India, 17–19 November 2016 , 1 – 6 .

Liu Z , Li N , Zhang C. Unit commitment scheduling using a hybrid ANN and Lagrangian relaxation method . In: 2008 International Conference on Multimedia and Ubiquitous Engineering (mue 2008), Busan, South Korea, 24–25 April 2008 , 481 – 484 .

Kumar VS , Mohan MR. Solution to security constrained unit commitment problem using genetic algorithm . International Journal of Electrical Power & Energy Systems , 2010 , 32 : 117 – 125 .

Prakash S , Sinha SK. Application of artificial intelligence in load frequency control of interconnected power system . International Journal of Engineering, Science and Technology , 2011 , 3 : 264 – 275 .

Moshtagh J , Rafinia A. A new approach to high impedance fault location in three-phase underground distribution system using combination of fuzzy logic & wavelet analysis . In: 2012 11th International Conference on Environment and Electrical Engineering, Venice, Italy, 15–18 May 2012 , 90 – 97 .

Santoso NT , Tan OT. Neural-net based realtime control of capacitors installed on distribution systems . IEEE Trans. Power Delivery , 1990 , 5 : 266 – 272 .

Ng HN , Salama MMA , Chikhani AY. Capacitor allocation by approximate reasoning: fuzzy capacitor placement . IEEE Trans Power Deliv , 2000 , 15 : 393 – 398 .

Zhang H , Zhang L , Meng F. Reactive power optimization based on genetic algorithm . In: POWERCON’98. International Conference on Power System Technology. Proceedings, Vol. 2 , Beijing, China, 18–21 August 1998 , 1448 – 1453 .

Mamandur KRC , Chenoweth RD. Optimal control of reactive power flow for improvements in voltage profiles and for real power loss minimization . IEEE Transactions on Power Apparatus and Systems, 1981 , PAS-100 : 3185 – 3194 .

Kothakotla V , Kumar B. Analysis and design of robust PID controller for grid voltage control of islanded microgrid using genetic algorithm . In: 2021 5th International Conference on Intelligent Computing and Control Systems (ICICCS), Madurai, India, 6–8 May 2021 , 719 – 726 .

Wang S , Duan J , Shi D , et al.  . A data-driven multi-agent autonomous voltage control framework using deep reinforcement learning . IEEE Trans Power Syst , 2020 , 35 : 4644 – 4654 .

Zidani Y , Zouggar S , Elbacha A. Steady-state analysis and voltage control of the self-excited induction generator using artificial neural network and an active filter . Iranian Journal of Science and Technology, Transactions of Electrical Engineering , 2018 , 42 : 41 – 48 .

Sumathi S. Artificial neural network application for voltage control and power flow control in power systems with UPFC . In: 2015 International Conference on Emerging Research in Electronics, Computer Science and Technology (ICERECT), Mandya, India, 17–19 December 2015 , 403 – 407 .

Kanata S , Sianipar GH , Maulidevi NU. Optimization of reactive power and voltage control in power system using hybrid artificial neural network and particle swarm optimization . In: 2018 2nd International Conference on Applied Electromagnetic Technology (AEMT), Lombok, Indonesia, 9–12 April 2018 , 67 – 72 .

Abdalla OH , Ghany AA , Fayek HH. Coordinated PID secondary voltage control of a power system based on genetic algorithm . In: 2016 Eighteenth International Middle East Power Systems Conference (MEPCON), Cairo, Egypt, 27–29 December 2016 , 214 – 219 .

Chung IY , Liu W , Cartes DA. Control parameter optimization for a microgrid system using particle swarm optimization . In: 2008 IEEE International Conference on Sustainable Energy Technologies, Singapore, 24–27 November 2008 , 837 – 842 .

Yousuf H , Zainal AY , Alshurideh M , et al.  . Artificial intelligence models in power system analysis . In: Hassanien AE , Bhatnagar R , Darwish A (eds). Artificial Intelligence for Sustainable Development: Theory, Practice and Future Applications . Cham, Switzerland : Springer , 2021 , 231 – 242 .

Aakula JL , Khanduri A , Sharma A. Determining reactive power levels to improve bus voltages using PSO . In: 2020 IEEE 17th India Council International Conference (INDICON), New Delhi, India, 10–13 December 2020 , 1 – 7 .

Karthikeyan R , Pasam S , Sudheer S , et al.  . Fuzzy fractional order PID based parallel cascade control system . In: Thampi S , Abraham A , Pal S , Rodriguez J (eds). Recent Advances in Intelligent Informatics. Advances in Intelligent Systems and Computing , Vol. 235 . Cham, Switzerland : Springer , 2014 , 293 – 302 .

Sallama A , Abbod M , Taylor G. Supervisory power system stability control using neuro-fuzzy system and particle swarm optimization algorithm . In: 2014 49th International Universities Power Engineering Conference (UPEC), Cluj-Napoca, Romania, 2–5 September 2014 , 1 – 6 .

Chen B , Li W , Yu P , et al.  . Planning approach for cross-regional optical transmission network supporting wide-area stability control services in power system . In: 2018 International Conference on Network Infrastructure and Digital Content (IC-NIDC), Guiyang, China, 22–24 August 2018 , 105 – 109 .

Torkzadeh R , NasrAzadani H , Aliabad AD , et al.  . A genetic algorithm optimized fuzzy logic controller for UPFC in order to damp of low frequency oscillations in power systems . In: 2014 22nd Iranian Conference on Electrical Engineering (ICEE), Tehran, Iran, 20–22 May 2014 , 706 – 712 .

Dutta S , Singh SP. Optimal rescheduling of generators for congestion management based on particle swarm optimization . IEEE Trans Power Syst , 2008 , 23 : 1560 – 1569 .

Nam S , Hur J. Probabilistic forecasting model of solar power outputs based on the naive Bayes classifier and kriging models . Energies , 2018 , 11 : 2982 .

Miraftabzadeh SM , Longo M , Foiadelli F , et al.  . Advances in the application of machine learning techniques for power system analytics: a survey . Energies , 2021 , 14 : 4776 .

Safari A , Babaei F , Farrokhifar M. A load frequency control using a PSO-based ANN for micro-grids in the presence of electric vehicles . Int J Ambient Energy , 2021 , 42 : 688 – 700 .

Joshi M , Sharma G , Davidson IE. Load frequency control of hydro electric system using application of fuzzy with particle swarm optimization algorithm . In: 2020 International Conference on Artificial Intelligence, Big Data, Computing and Data Communication Systems (icABCD), Durban, South Africa, 6–7 August 2020 , 1 – 6 .

Nguyen GN , Jagatheesan K , Ashour AS , et al.  . Ant colony optimization-based load frequency control of multi-area interconnected thermal power system with governor dead-band nonlinearity . In: Yang XS , Nagar AK , Joshi A (eds). Smart Trends in Systems, Security and Sustainability . Singapore : Springer , 2018 , 157 – 167 .

Balamurugan CR. Three area power system load frequency control using fuzzy logic controller . International Journal of Applied Power Engineering (IJAPE) , 2018 , 7 : 18 – 26 .

Otani T , Tanabe R , Koyanagi Y , et al.  . Cooperative load frequency control of generator and battery using a recurrent neural network . In: TENCON 2017–2017 IEEE Region 10 Conference, Penang, Malaysia, 5–8 November 2017 , 918 – 923 .

Kumar D , Mathur HD , Bhanot S , et al.  . Forecasting of solar and wind power using LSTM RNN for load frequency control in isolated microgrid . Int J Modelling Simul , 2021 , 41 : 311 – 323 .

Arora K , Kumar A , Kamboj VK , et al.  . Optimization methodologies and testing on standard benchmark functions of load frequency control for interconnected multi area power system in smart grids . Mathematics , 2020 , 8 : 980 .

Vadivelu KR , Marutheswar GV. Artificial intelligence technique based reactive power planning incorporating FACTS controllers in real time power transmission system . In: 2014 IEEE 2nd International Conference on Electrical Energy Systems (ICEES), Chennai, India, 7–9 January 2014 , 26 – 31 .

Khan MA , Algarni F. A healthcare monitoring system for the diagnosis of heart disease in the IoMT cloud environment using MSSO-ANFIS . IEEE Access , 2020 , 8 : 122259 – 122269 .

Liu M , Dong M , Wu C. A new ANFIS for parameter prediction with numeric and categorical inputs . IEEE Trans Autom Sci Eng , 2010 , 7 : 645 – 653 .

Ghadiri A , Haghifam MR , Larimi SMM. Comprehensive approach for hybrid AC/DC distribution network planning using genetic algorithm . IET Generation, Transmission & Distribution , 2017 , 11 : 3892 – 3902 .

Jain SK , Bhargava A , Pal RK. Three area power system load frequency control using fuzzy logic controller . In: 2015 International Conference on Computer, Communication and Control (IC4), Indore, India, 10–12 September 2015 , 1 – 6 .

Son YS , Kim HJ , Kim JT. A video-quality control scheme using ANFIS architecture in a DASH environment . Journal of Broadcast Engineering , 2018 , 23 : 104 – 114 .

Kannadasan K , Edla DR , Yadav MH , et al.  . Intelligent-ANFIS model for predicting measurement of surface roughness and geometric tolerances in three-axis CNC milling . IEEE Trans Instrum Meas , 2020 , 69 : 7683 – 7694 .

Penghui L , Ewees AA , Beyaztas BH , et al.  . Metaheuristic optimization algorithms hybridized with artificial intelligence model for soil temperature prediction: novel model . IEEE Access , 2020 , 8 : 51884 – 51904 .

Tawfeek TS , Ahmed AH , Hasan S. Analytical and particle swarm optimization algorithms for optimal allocation of four different distributed generation types in radial distribution networks . Energy Procedia , 2018 , 153 : 86 – 94 .

Yeom CU , Kwak KC. Adaptive neuro-fuzzy inference system predictor with an incremental tree structure based on a context-based fuzzy clustering approach . Applied Sciences , 2020 , 10 : 8495 .

Krasopoulos CT , Beniakar ME , Kladas AG. Multicriteria PM motor design based on ANFIS evaluation of EV driving cycle efficiency . IEEE Trans Transp Electrif , 2018 , 42 : 525 – 535 .

El-Hasnony IM , Barakat SI , Mostafa . Optimized ANFIS model using hybrid metaheuristic algorithms for Parkinson’s disease prediction in IoT environment . IEEE Access , 2020 , 8 : 119252 – 119270 .

Morshedizadeh M , Kordestani M , Carriveau R , et al.  . Power production prediction of wind turbines using a fusion of MLP and ANFIS networks . IET Renew Power Gener , 2018 , 12 : 1025 – 1033 .

Khosravi A , Nahavandi S , Creighton D. Prediction interval construction and optimization for adaptive neurofuzzy inference systems . IEEE Trans Fuzzy Syst , 2011 , 19 : 983 – 988 .

Elbaz K , Shen SL , Sun WJ , et al.  . Prediction model of shield performance during tunneling via incorporating improved particle swarm optimization into ANFIS . IEEE Access , 2020 , 8 : 39659 – 39671 .

Dovžan D , Škrjanc I. Fuzzy space partitioning based on hyperplanes defined by eigenvectors for Takagi-Sugeno fuzzy model identification . IEEE Trans Ind Electron , 2019 , 67 : 5144 – 5153 .

Castiello C , Fanelli AM , Lucarelli M , et al.  . Interpretable fuzzy partitioning of classified data with variable granularity . Appl Soft Comput , 2019 , 74 : 567 – 582 .

Alexandridis A , Chondrodima E , Sarimveis H. Radial basis function network training using a nonsymmetric partition of the input space and particle swarm optimization . IEEE Trans Neural Networks Learn Syst , 2012 , 24 : 219 – 230 .

Verstraete J. The spatial disaggregation problem: simulating reasoning using a fuzzy inference system . IEEE Trans Fuzzy Syst , 2016 , 25 : 627 – 641 .

Olamaei J , Moradi M , Kaboodi T. A new adaptive modified firefly algorithm to solve optimal capacitor placement problem . In: 18th Electric Power Distribution Conference, Kermanshah, Iran, 30 April–1 May 2013 , 1 – 6 .

Lee JS , Teng CL. An enhanced hierarchical clustering approach for mobile sensor networks using fuzzy inference systems . IEEE Internet Things J , 2017 , 4 : 1095 – 1103 .

Su ZG , Denoeux T. BPEC. Belief-peaks evidential clustering . IEEE Trans Fuzzy Syst , 2018 , 27 : 111 – 123 .

Xu P , Deng Z , Cui C , et al.  . Concise fuzzy system modeling integrating soft subspace clustering and sparse learning . IEEE Trans Fuzzy Syst , 2019 , 27 : 2176 – 2189 .

Sujil A , Kumar R , Bansal RC. FCM clustering-ANFIS-based PV and wind generation forecasting agent for energy management in a smart microgrid . The Journal of Engineering , 2019 , 2019 : 4852 – 4857 .

Neamatollahi P , Naghibzadeh M , Abrishami S. Fuzzy-based clustering-task scheduling for lifetime enhancement in wireless sensor networks . IEEE Sens J , 2017 , 17 : 6837 – 6844 .

Wang Z , Xu Y. A golden section-based double population genetic algorithm applied to reactive power optimization . IOP Conference Series: Earth and Environmental Science, 2021 , 645 : 012074 .

Kahouli O , Alsaif H , Bouteraa Y , et al.  . Power system reconfiguration in distribution network for improving reliability using genetic algorithm and particle swarm optimization . Applied Sciences , 2021 , 11 : 3092 .

Žarković SD , Stanković S , Shayesteh E , et al.  . Reliability improvement of distribution system through distribution system planning: MILP vs. GA . In: 2019 IEEE Milan PowerTech, Milan, Italy, 23–27 June 2019, 1 – 6 .

Ahmetovic H , Saric M , Hivziefendic J. Reliability based power distribution network planning using fuzzy logic . Advances in Electrical and Electronic Engineering , 2021 , 19 : 123 – 133 .

Suresh M , Sirish TS , Subhashini TV , et al.  . Load flow analysis of distribution system using artificial neural networks . In: Satapathy SC , Bhateja V , Udgata SK , Pattnaik PK (eds). Proceedings of the 5th International Conference on Frontiers in Intelligent Computing: Theory and Applications FICTA 2016, Vol. 1 . Springer , Singapore , 2017 , 515 – 524 .

Kumari M , Ranjan R , Singh VR. Optimal power distribution planning using improved particle swarm optimization . International Journal of Intelligent Systems and Applications in Engineering , 2018 , 3 : 170 – 177 .

Saha S , Mukherjee V. A novel multi-objective modified symbiotic organisms search algorithm for optimal allocation of distributed generation in radial distribution system . Neural Comput Appl , 2021 , 33 : 1751 – 1771 .

Hosseini MM , Parvania M. Artificial intelligence for resilience enhancement of power distribution systems . The Electricity Journal , 2021 , 34 : 106880 .

Lytras MD , Chui KT. The recent development of artificial intelligence for smart and sustainable energy systems and applications . Energies , 2019 , 12 : 3108 .

Gandhar S , Ohri J , Singh M. Dynamic reactive power optimization of hybrid micro grid in islanded mode using fuzzy tuned UPFC . J Inf Optim Sci , 2020 , 41 : 305 – 315 .

Tang Y , Hu W , Cao D , et al.  . Artificial intelligence-aided minimum reactive power control for the DAB converter based on harmonic analysis method . IEEE Trans Power Electron , 2021 , 36 : 9704 – 9710 .

Harrye YA , Ahmed KH , Aboushady AA. Reactive power minimization of dual active bridge DC/DC converter with triple phase shift control using neural network . In: 2014 International Conference on Renewable Energy Research and Application (ICRERA), Milwaukee, WI, USA, 19–22 October 2014 , 566 – 571 .

Wang Y , Li Y. A hybrid ant colony optimization algorithm for dynamic optimization of vehicle routing problem with time windows . Applied Intelligence , 2021 , 51 : 476 – 491 .

Sharma NK , Babu DS , Choube SC. Application of particle swarm optimization technique for reactive power optimization . In: IEEE International Conference on Advances in Engineering, Science and Management (ICAESM-2012), Nagapattinam, India, 30–31 March 2020 , 88 – 93 .

Bhattacharyya B , Gupta VK. Fuzzy based evolutionary algorithm for reactive power optimization with FACTS devices . International Journal of Electrical Power & Energy Systems , 2014 , 61 : 39 – 47 .

Bharti D. Multi-point optimal placement of shunt capacitor in radial distribution network: a comparison . In: 2020 International Conference on Emerging Frontiers in Electrical and Electronic Technologies (ICEFEET), Patna, India, 10–11 July 2020 , 1 – 6 .

Roy PK , Sultana S. Optimal reconfiguration of capacitor based radial distribution system using chaotic quasi oppositional chemical reaction optimization . Microsyst Technol , 2022 , 28 : 499 – 511 .

Pimentel Filho MC , De Lacerda EGM , Junior MM. Capacitor placement using ant colony optimization and gradient . In: 2009 15th International Conference on Intelligent System Applications to Power Systems, Curitiba, Brazil, 8–12 November 2009 , 1 – 4 .

Isac SJ , Kumar KS , Kumar PV. Optimal capacitor placement in radial distribution system to minimize the loss using fuzzy logic control . In: International Conference on Smart Structures and Systems—ICSSS; 13 , Chennai, India, 28–29 March 2013 , 33 – 40 .

Reddy MD , Reddy VV. Capacitor placement using fuzzy and particle swarm optimization method for maximum annual savings . ARPN Journal of Engineering and Applied Sciences , 2008 , 3 : 25 – 30 .

Shwehdi MH , Mohamed SR , Devaraj D. Optimal capacitor placement on West–East inter-tie in Saudi Arabia using genetic algorithm . Computers & Electrical Engineering , 2018 , 68 : 156 – 169 .

Mahdavian M , Kafi MH , Movahedi A , et al.  . Improve performance in electrical power distribution system by optimal capacitor placement using genetic algorithm . In: 2017 14th International Conference on Electrical Engineering/Electronics, Computer, Telecommunications, and Information Technology (ECTI-CON), Phuket, Thailand, 27–30 June 2017 , 749 – 752 .

Email alerts

Citing articles via.

• Advertising and Corporate Services

Affiliations

• Online ISSN 2515-396X
• Print ISSN 2515-4230
• Copyright © 2024 National Institute of Clean-and-Low-Carbon Energy
• About Oxford Academic
• Publish journals with us
• University press partners
• What we publish
• New features  
• Open access
• Institutional account management
• Rights and permissions
• Get help with access
• Accessibility
• Advertising
• Media enquiries
• Oxford University Press
• Oxford Languages
• University of Oxford

Oxford University Press is a department of the University of Oxford. It furthers the University's objective of excellence in research, scholarship, and education by publishing worldwide

• Copyright © 2024 Oxford University Press
• Cookie settings
• Cookie policy
• Privacy policy
• Legal notice

This Feature Is Available To Subscribers Only

Sign In or Create an Account

This PDF is available to Subscribers Only

For full access to this pdf, sign in to an existing account, or purchase an annual subscription.

Recent Trends In Power System Protection

First Page of the Article

IEEE Account

• Change Username/Password
• Update Address

Purchase Details

• Payment Options
• Order History
• View Purchased Documents

Profile Information

• Communications Preferences
• Profession and Education
• Technical Interests
• US & Canada: +1 800 678 4333
• Worldwide: +1 732 981 0060
• Contact & Support
• About IEEE Xplore
• Accessibility
• Terms of Use
• Nondiscrimination Policy
• Privacy & Opting Out of Cookies

A not-for-profit organization, IEEE is the world's largest technical professional organization dedicated to advancing technology for the benefit of humanity. © Copyright 2024 IEEE - All rights reserved. Use of this web site signifies your agreement to the terms and conditions.

Click through the PLOS taxonomy to find articles in your field.

For more information about PLOS Subject Areas, click here .

Loading metrics

Open Access

Peer-reviewed

Research Article

Relay protection system of transmission line based on AI

Roles Writing – review & editing

* E-mail: [email protected]

Affiliations School of Electrical Engineering, Xi’an University of Technology, Xi’an, Shaanxi, China, State Grid GanSu Electric Power Research Institution, Lanzhou, Gansu, China

Roles Writing – original draft

Affiliations School of Electrical Engineering, Xi’an University of Technology, Xi’an, Shaanxi, China, Energy Intelligence Laboratory, Xi’an University of Technology, Xi’an, Shaanxi, China

Roles Visualization

Affiliation Lanzhou Petrochemical College of Vocational Technology, Lanzhou, Gansu, China

Roles Software

Affiliation Xinjiang Goldwind Science & Technology Co., Ltd, Urumqi, Xinjiang, China

Roles Supervision

Affiliation State Grid GanSu Electric Power Research Institution, Lanzhou, Gansu, China

Roles Methodology

Affiliation School of Electrical Engineering, Xi’an University of Technology, Xi’an, Shaanxi, China

• Xiangyu Zheng, 
• Rong Jia, 
• Linling Gong, 
• Aisikaer, 
• Xiping Ma, 

• Published: April 7, 2021
• https://doi.org/10.1371/journal.pone.0246403
• Reader Comments

With the development of modern power systems, higher requirements are imposed on relay protection technology. Traditional relay protection and fault diagnosis technologies have been unable to meet the requirements of the continuous development of power systems, and relay protection systems based on artificial intelligence(AI) technology have received increasing attention. Therefore, this document first analyses the weaknesses of traditional broadcast line protection and uses the adaptability and self-learning of artificial intelligence(AI); to propose the concept of protection of a relay line based on AI. In combination with the artificial nervous network, the AI-based relay protection system shall be studied and the experimental model shall be developed. This paper validates it with simulation experiments. The research results show that for the analysis of the ANN test results of the subnetwork, the actual output of the subnetwork is very close to the ideal output, and the error does not exceed 0.2%. The system has good performance and high reliability.

Citation: Zheng X, Jia R, Gong L, Aisikaer, Ma X, Dang J (2021) Relay protection system of transmission line based on AI. PLoS ONE 16(4): e0246403. https://doi.org/10.1371/journal.pone.0246403

Editor: Zhihan Lv, University College London, UNITED KINGDOM

Received: September 24, 2020; Accepted: January 19, 2021; Published: April 7, 2021

Copyright: © 2021 Zheng et al. This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

Data Availability: All relevant data are within the manuscript and its Supporting Information files.

Funding: This study was supported in the form of funding by Key project of Natural Science Basic Research plan in Shaanxi Province of China (Grant No. 2018ZDXM-GY-169) and Key project of Natural Science Basic Research plan in Shaanxi Province of China (Grant No. 2019ZDLGY18-03) for authors XZ, RJ, and JD. Xinjiang Goldwind Technology Co., Ltd. provided funding via salary for SA; and State Grid Gansu Electric Power Research Institute provided funding via salary for XM. The specific roles of these authors are articulated in the ‘author contributions’ section. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Competing interests: The authors have read the journal’s policy and the authors of this manuscript have the following competing interests: SA is a paid employee of Xinjiang Goldwind Technology Co., Ltd. and XM is a paid employee of State Grid Gansu Electric Power Research Institute There are no patents, products in development or marketed products associated with this research to declare. This does not alter our adherence to PLOS ONE policies on sharing data and materials.

Introduction

In daily life, many people have applied robots to various aspects. Development has brought convenience to people’s lives. During the operation of the power system, various faults and abnormal operating conditions may occur. The most common and most dangerous faults are various types of short circuits, including phase-to-phase shorts and ground shorts. The occurrence of system faults and abnormal operating conditions is inevitable. Once a fault occurs, it will affect other non-faulty equipment and even cause new faults. In order to prevent system accidents from expanding, to ensure that non-faulty parts can still reliably supply power, and to maintain the stability of power system operation, it is required to quickly and selectively remove faulty components. The time for removing the fault is extremely short. Obviously, it is impossible for the operating personnel to find the faulty equipment and remove the faulty equipment. Therefore, it is necessary to rely on an automatic device installed on each electrical equipment, that is, a relay protection device, to achieve it accurately.

At present, a relay protection device that responds to power frequency electrical quantities is widely used on power transmission lines. With the rapid development of the power system, the emergence and increase of large-capacity units and ultra-high-voltage transmission lines have put forward higher requirements for the reliability and speed of the relay protection operation. For a long time, in order to meet the needs of power systems, people have continuously tried to combine relay protection and fault location. However, there are many unsolvable contradictions in relay protection and fault location based on power frequency. The principle has important theoretical and practical significance. The neural network in AI has parallel computing capabilities, strong adaptability, high robustness and fault tolerance. It has unparalleled advantages over traditional computing methods for solving nonlinear systems, and it makes up for the simplicity of traditional methods. Relying on the lack of mathematical solutions, it solves some problems that are difficult or impossible to solve with traditional calculation methods. There are many extremely complex engineering calculations and nonlinear optimization problems. Agent is practical in the field of power control, and it may be applied to the field of relay protection. The introduction of Agent into the field of relay protection, combined with the nervous system, is of great significance for solving the problem of relay protection of transmission lines.

Stanislav Misak has summarized several operation modes in the past ten years, and Stanislav Misak has proposed transmission line protection (TLP) based on distance relay and TLP distance relay model based on analog anti-aliasing filter. The transmission line protection system is mainly composed of a circuit breaker (CB), a distance relay (DR) and an overcurrent relay (OCR). According to the medium length (10km-200km) of the transmission line (TL), distance relays are used for zone protection [ 1 , 2 ].

TU Qingrui once proposed a transmission line and differential impedance relay. The internal and external faults have different impedances. The working area of the sum impedance relay is set on the impedance plane [ 3 , 4 ]. With the trend of digitalization and intelligentization of substations, the dependence on the relay protection time synchronization system is increasing. When the time synchronization system breaks down, not only can’t phase comparison be done by the protection of the power on both sides of the line, even the protection of local power based on impedance / direction information will fail. In this case, the existing protection principle will fail, and only the fast-breaking current protection and remote backup protection of other substations can be relied on, and the scope and speed of protection will be seriously deteriorated. In response to this problem, on the basis of high-resistance ground fault detection capability and resistance protection that does not require time synchronization, Jakub Jedrzejczak once proposed a mode and impedance protection criterion without phase information synchronization, combined with quick-break overcurrent protection and half-mode protection have verified the validity of this criterion through simulation tests [ 5 , 6 ].

The innovation of this paper lies in introducing Agent into the field of relay protection, combining it with the nervous system, and establishing a simulation model to simulate and solve the problem of relay protection of transmission lines, and verify the reliability of the method proposed in this paper.

Relay protection method for transmission line based on AI

Current protection.

(1) Working principle of current protection

According to the requirements of line faults for main and backup protection, there are three types of current protection for transmission lines:

• Untimed current instantaneous trip protection, referred to as the first stage of current protection. Its role is to ensure that only faults on this line are removed under any circumstances.
• The setting value of its current measurement element must follow the following principles: The time-limit current quick-break protection can protect the entire length of the line (including the end of the line). To this end, the protection range must be extended to the adjacent lower ~ line.
• Time-limit overcurrent protection, referred to as the third stage of current protection, which is used as the backup of the main protection of this line ~ the backup protection of the line (or components), that is, the remote backup protection. The starting current of the current protection is to avoid the maximum load. Sections 1, 2, and 3 are collectively referred to as three-phase current protection for line short circuits. In the single-power radiation network, the time-limited current quick-break protection at the circuit breaker 1QF is shown in Fig 1 .

• PPT PowerPoint slide
• PNG larger image
• TIFF original image

https://doi.org/10.1371/journal.pone.0246403.g001

(2) Advantages and disadvantages of current protection

In the case where the system operation mode varies greatly, when the 1QF current quick-break protection of the circuit breaker is set according to the selective requirements of protection under the maximum operation mode, there is no protection range under the minimum operation mode.

Distance protection

The current protection of transmission lines has a simple structure and good reliability, and is used for the requirements of medium and low voltage electrical protection performance.

(1) Working principle of distance protection

The working principle of distance protection is shown in Fig 2 .

https://doi.org/10.1371/journal.pone.0246403.g002

(2) Advantages and disadvantages of distance protection

Its main advantages and disadvantages are as follows:

• In multi-supply networks and even in complex power networks, distance protection can better meet the selective requirements of the actions.
• The first stage of the protection distance is protection for instantaneous action to limit the damage to the first protection zone. In a dual-supply network, if the first-stage protection on both sides of the line has overlapping protection zones, both sides of the line may remove errors in the overlay zone without delay, and the radiation network of a source in the first stage protection zone of the line after the first step error and error non-overlapping area on both sides of the dual supply line, the action shall not be removed without delay.
• The wiring of the components of composite resistance to distance protection is more complex and the relevant locking device must be added to make the device distance more complicated. Therefore, the reliability of the distance protection is lower than that of the current protection.

Artificial neural networks

(1) Artificial neuron model

The most typical artificial neuron model is shown in Fig 3 .

https://doi.org/10.1371/journal.pone.0246403.g003

Among them, θ j is called threshold, w ji is called connection weight coefficient, and f () is called output transformation function.

(2) Artificial neural network model

Strictly speaking, a neural network is a directed graph with the following properties.

• There is a state variable x j for each node;
• There is a connection weight coefficient w ji from node i to node j ;
• There is a threshold θ j for each node;

A neural network model is represented by network topology, node characteristics, and learning rules. Artificial neural networks are a simulation and approximation of biological neural networks.

(3) Characteristics of artificial neural network

The neural networks have parallel large-scale processing and distributed data storage capabilities and have good self-adaptation, self-organisation and strong learning function, correlation function and error tolerance [ 7 , 8 ].

1) With an adaptive function

It is mainly based on the data provided, through learning and training, to discover the internal relationship between exit and exit, so as to find a solution to the problem, Lee is not based on previous knowledge and rules of the problem, so he has a good Sex [ 9 , 10 ].

2) With generalization function

It can process untrained data and obtain appropriate answers corresponding to these data. Similarly, it is able to handle noisy or incomplete data, thus showing good fault tolerance [ 11 , 12 ].

3) Non-linear mapping function

The actual problems are often very complicated, and the factors affect each other, showing a complex nonlinear relationship. Artificial neural network is a kind of non-linear processing unit. Neural network provides useful tools for dealing with these problems [ 13 , 14 ].

4) Highly parallel processing

The processing of neural network can be highly parallel, so the processing speed of neural network realized by hardware can be much higher than that of ordinary computer [ 15 , 16 ].

Multi-agent system

The multi-agency system is the pioneering field of AI today and an important area of decentralised AI research. The aim is to build large and complex systems (software and material systems) on a small scale, to communicate with each other, make simple adjustments and manage object systems [ 17 , 18 ].

(1) Definition and characteristics of Agent

Agent definition includes two sub definitions: weak definition and strong definition [ 19 , 20 ]. Strong definition thinks that Agent should include human characteristics such as emotion, belief and intention on the basis of weakly defined characteristics [ 21 , 22 ].

It is generally believed that the Agent has the following characteristics:

1) Autonomy

An Agent should have independent knowledge and knowledge processing methods that are local to it. With its own limited computing resources and behavior control mechanisms, it can continue to operate without direct intervention and guidance from humans and other Agents in a specific way. Respond to environmental requirements and changes.

2) Social ability or communication

Agents often do not exist independently. Just like biological groups in the real world, many agents often exist in the environment at the same time, forming a social group.

3) Responsiveness

That is, the perception and influence of the environment, whether the subject lives in the real world (such as robots, communication subjects on the Internet, user interface subjects, etc.).

4) Self-issued

Traditional applications are passively run by the user and mechanically complete the user’s instructions. The behavior of the intelligent agent should be active, or spontaneous, and the agent can sense changes in the surrounding environment and make goal-based behaviors.

(2) Model structure of Agent

The intelligence of the Agent is realized through its structure. There are various views on the model structure of the Agent. It is generally believed that an Agent should include sensors, decision controllers, mental states, knowledge bases, and communicators. The BDI model is a generally accepted Agent. The mental state model, which believes that the agent’s mental factors include Belief, Desire, and Intention. Belief represents the agent’s understanding of the objective world and is the basis for his judgment and decision-making; Desire describes what the agent wants to achieve; Intention is the action it takes to achieve its goal.

(3) The main types of Agent

Agent technology, as a new method of computer software development, is a relatively large category. From ordinary people, animals, social institutions, and even ordinary objects, they can be abstracted as Agents according to different applications [ 23 ]. On the system structure, according to the hierarchical model of human thinking, the internal system of the Agent can be divided into the following three categories:

1) Active Agent

An agent with an active structure is also called a cautious agent or a cognitive agent. It is an explicit symbolic model that includes logical reasoning capabilities for the environment and intelligent behavior. It maintains the tradition of classic AI and is a knowledge-based system. Among them, the environment model is generally realized in advance, forming the knowledge base of the main components. It is considered that the Agent has some limitations in adapting to the dynamic environment. Think carefully Agent is an active software with internal states. It is different from specific domain knowledge, and it has knowledge representation, problem solving representation, environment representation, and specific communication protocols. The Agent receives the information of the external environment through sensors, fuses information according to the internal state, generates a description to modify the current state, and then formulates a plan with the support of the knowledge base to form a series of actions that affect the environment through effectors.

2) Reactive Agent

Consider carefully that Agent’s symbolic algorithms are generally designed for ideal and provable results, often resulting in high complexity. In contrast, reactive agents are agents that do not include symbolic world models and do not use complex symbolic reasoning.

3) Hybrid Agent

It can be seen that the cautious agent has high intelligence, but has a slow response and low execution efficiency; while a reactive agent has a fast response and can adapt quickly to environmental changes, it has low intelligence and is not flexible enough. Combining the advantages of the two, forming a hybrid Agent has better flexibility and faster response speed. A hybrid agent consists of two (or more) subsystems in an agent: one is a careful thinking subsystem, which contains a world model represented by symbols, and uses mainstream AI to generate plans and decisions; the other is a reaction subsystem, used to handle events without complex reasoning.

4) Multi-Agent System

A multi-agent system is a collection of multiple interacting and interconnected agentsAgents, reality systems mostly belong to multi-agent systems [ 24 , 25 ].

Transmission line protection system based on multi-agent technology

(1) Features and structure of MAS-based transmission line protection systems

The characteristics of the MAS-based protection system are as follows:

• Multiple agents are distributed hierarchically, corresponding to different levels of tasks.
• Each Agent has its own goals and can complete its own tasks autonomously.
• Agent can generate new task agents, activate superior agents or interact with agents at the same level as needed.
• Each Agent can communicate with each other and share information through communication. Can use distributed multi-point information to determine faults, speed up protection actions, and remove faults.
• Multiple agents can cooperate with each other, and coordinate the action of multiple agents in the same layer and different layers, thereby improving the protection adaptability.
• A protection principle can be completed by a combination of multiple agents, which is flexible and portable.

The Agent in the protection system can be large or small. It can be a simple detection link, or it can be different protection devices in the substation. Their commonality is based on cooperation.

(2) Decision-making process of protection system

The organic coating factor can be adjusted to the S3 stack substrate and other substrates communicate with substation s3 via fibre optics. Within the Agent Agency, according to the network wiring, corresponding circuit switches have been stored to form the same protection field. When the organisation agent receives the action message interrupted by the circuit breaker, quickly considers that other switches within the protection range should be activated and transmits this directive to the protection subsystems on all sides through the factor; coordination. Any protection agent may use the current protection on the basis of a neural network consisting of two neural networks. The neural network 1 (ANN 1) is a positive-direction separator and a failure type selector on adjacent lines. The neural network 2 (ANN 2) is a sub-network for instantaneous action when an error occurs within 85% of the protection zone. When sub-network 1 and sub-network 2 operate simultaneously, the travel agent shall be activated and the local protection switch shall be activated instantaneously; and that information shall be transmitted to the Agency Agent at the same time. When Sub-Network 1 operates and receives the travel order transmitted by the Agency Agent, the travel agent shall also activate and immediately activate the local protection switch.

Experimental verification

Experimental design.

Four independent neural sub-networks are designed, which are direction discrimination sub-network ANN 1, fault type and phase discrimination sub-network ANN 2, instantaneous exit action sub-network ANN 3, and adjacent circuit breaker trip detection sub-network ANN 4.

(1) Direction discrimination sub-network ANN l

The role of the direction discrimination sub-network is to determine the direction in which a short-circuit fault occurs, that is, to distinguish between a short-circuit fault in the forward direction or a short-circuit fault in the reverse direction. According to the above analysis, when short-circuit faults occur in different directions, the phase angle between voltage and current will change significantly. In order to prevent the saturation of the neuron nodes, the phase angle is divided by Ⅱ, and the input feature quantity of ANN l is limited to [–1, 1]. Select the direction to specify the exit node of the infrastructure is 1, the output value is 1 when the forward shortcut appears and the output value is 0 when the short circuit reversal error appears. The neural net 1 training samples are shown in Table 1 .

https://doi.org/10.1371/journal.pone.0246403.t001

(2) ANN 2 for fault type and phase discrimination sub-network

The function of fault type and phase discrimination sub-network is to determine the type of short-circuit fault (single-phase grounding or phase-to-phase short-circuit) and the fault phase difference, so as to ensure that various faults can be correctly judged within the protection scope. According to the above analysis, when a short-circuit fault occurs, the phase current of the fault phase increases sharply. Therefore, the amplitude of the three-phase fundamental wave current before and after the fault is selected as the characteristic input of ANN2. There are 6 input nodes in total. In order to speed up the protection operation speed, a half-wave Fourier algorithm is used to obtain the amplitude of the fundamental wave currents of each phase, and the maximum short-circuit current flowing at the protection installation is used as a reference value. There are 4 output nodes of the fault type and phase judgment.

(3) Instantaneous exit action sub-network ANN 3

The role of the instantaneous exit action sub-network is to issue an instantaneous exit trip signal when it is determined that a short-circuit fault occurs within 85% of the total length of the line. The selection of its input feature is similar to ANN2, except that the range of ANN2 extends to 50% of the adjacent line, and the range of ANN3 is only 85% of the line. The output node of the instantaneous exit action sub-network is one. The output value of the short-circuit fault is 1 within 85% of the total length of the line, and the output value is 0 when the short-circuit fault exceeds 85% of the total length of the line.

(4) ANN 4 on the opposite side

The role of the opposite circuit breaker trip detection sub-network is to determine whether the circuit breaker on the opposite side of the line is tripped. When a short-circuit fault occurs on the line, the protection on the side close to the fault point will instantly act on the circuit breaker trip. If it is an out-of-area short-circuit fault, when the adjacent line circuit breaker trips, the fault point is isolated, the system resumes normal operation, and the line current suddenly changes from the fault current to the load current. After the circuit breaker trips, the fault point still exists, and the fault current will continue to flow on the line.

Analysis of results

Analysis of sample inspection results.

(1) Training and testing of ANN 1

Select a sample different from the training sample as a test sample to test the trained neural network. Part of the training and test results are shown in Table 2 and Fig 4 .

https://doi.org/10.1371/journal.pone.0246403.g004

https://doi.org/10.1371/journal.pone.0246403.t002

It can be seen from Table 2 and Fig 4 that the actual output of the sub-network is very close to the ideal output. For faults at different locations of the same fault type, the phase angle between voltage and current is the same. Therefore, the training samples in the forward direction are selected with different fault types and different transition resistances at the end of the line and adjacent lines at 50%, and the training samples in the reverse direction are also selected with different fault types and different transition resistances.

(2) Training and testing of ANN 2

The purpose of this sub-network is to identify the types and differences of faults, so it is necessary to consider various possible situations. In order to test the training results, a test sample different from the training sample is selected for testing. Partial training and test results are shown in Fig 5 .

https://doi.org/10.1371/journal.pone.0246403.g005

As can be seen from the figure, the actual output of the sub-network is very close to the ideal output. The selected training samples include various short-circuit faults at different locations (different locations on this line and different locations within 50% of adjacent lines), with phase A metal grounding, and phase A passing different transition resistances (for 110kv networks, consider the maximum transition resistance) 50Q) Grounding, phase B metal grounding, phase B grounding through different transition resistances, phase C metal grounding, phase c grounding through different transition resistances, AB phase short circuit, AB phase short circuit ground, BC phase short circuit, BC phase short circuit ground, CA phase short circuit, CA phase break short circuit to ground, ABC three phase short circuit, etc.

(3) Training and testing of the sub-network ANN 3

The purpose of this sub-network is to identify the scope of the fault, so it is necessary to consider various possible situations. In order to test the training results, a test sample different from the training sample is selected for testing. Part of the training and test results are shown in Fig 6 .

https://doi.org/10.1371/journal.pone.0246403.g006

It can be seen from Fig 6 that the actual output of the sub-network is very close to the ideal output. The selected training samples include various short-circuit faults at different locations on this line, with phase A metal grounding, phase A grounding through different transition resistances (for a 110kV network, the maximum transition resistance is considered 50Q), phase B metal grounding, and phase B passages different transition resistance grounding, phase C metal grounding, phase C grounding through different transition resistances, AB phase short circuit, AB phase short circuit ground, BC phase short circuit, BC phase short circuit ground, CA phase short circuit, CA phase short circuit ground, ABC three phase short circuit Wait.

(4) Analysis of test results of sub-network ANN4

The purpose of this sub-network is to identify what kind of state the circuit breaker on the opposite side of the protection presents after a system failure? Is it a closed position or an open position, it is necessary to consider various possible situations. In order to test the training results, a test sample different from the training sample is selected for testing. Partial training and testing results are shown in Fig 7 .

https://doi.org/10.1371/journal.pone.0246403.g007

It can be seen from Fig 7 that the actual output of the sub-network is very close to the ideal output. The selected training samples include various short-circuit faults occurring in the protection zone and outside the zone, including phase A metal grounding, phase A grounding with different transition resistances (for a 110kV network, the maximum transition resistance is considered 50Q), phase B metal grounding, and B Phase to ground through different transition resistances, Phase C metal grounding, Phase C to ground through different transition resistances, AB phase short circuit, AB phase short circuit ground, BC phase short circuit, BC phase short circuit ground, CA phase short circuit, CA phase short circuit ground, ABC Three-phase short circuit, etc.

From the comprehensive results, for the analysis of the ANN test results of the subnetwork, the actual output of the subnetwork is very close to the ideal output, and the error does not exceed 0.2%.

Comparison with configured distance protection

Two types of protection were simulated with EMTDC simulation software. The area covered by the protection range of section one is only 15 km. The time of action of the distance protection for different error points shall be shown in Table 3 and Fig 8 .

https://doi.org/10.1371/journal.pone.0246403.g008

https://doi.org/10.1371/journal.pone.0246403.t003

For the faults that occur in most sleeping areas, there are always circuit breakers on one or both sides. Delay is required to trip; for MAS protection system, if any point in the protection area fails, the circuit breakers on each side can instantaneously trip.

Conclusions

With the further development of computer science and artificial intelligence technology, people’s lives have undergone earth-shaking changes. Artificial intelligence has brought great convenience to people’s lives. Humans are gradually replacing humans with machines to do things that are inconvenient for humans and have achieved certain results.

The application of AI in the power system is very helpful to solve the problem of relay protection of transmission lines. On this basis, this article introduces the research of AI-based relay protection system. This paper proposes a multi-factor relay protection system based on the basic framework of the agent. Discuss the decision-making process of the protection system and simulate the cooperative relationship between participants.

Now, fibre optics have been communicated between each substation in the power system and between the substation and the end of the mission. Some areas have already established fibre ring networks. The high speed and reliability of fibre communication guarantee the implementation of the MAS protection. With the in-depth study of Agent technology, the MAS will be well developed and implemented in the field of relay protection.

• View Article
• Google Scholar

Preprint  

• Preprint wes-2024-61

Functional Specifications and Testing Requirements of Grid-Forming Offshore Wind Power Plants

Abstract. Throughout the past few years, various transmission system operators (TSOs) and research institutes have defined several functional specifications for grid-forming (GFM) converters via grid codes, white papers, and technical documents. These institutes and organisations also proposed testing requirements for general inverter-based resources (IBRs) and specific GFM converters. This paper initially reviews functional specifications and testing requirements from several sources to create an understanding of GFM capabilities in general. Furthermore, it proposes an outlook of the defined GFM capabilities, functional specifications, and testing requirements for offshore wind power plant (OF WPP) applications from an original equipment manufacturer (OEM) perspective. Finally, this paper briefly establishes the relevance of new testing methodologies for equipment-level certification and model validation, focusing on GFM functional specifications.

• Preprint (PDF, 1514 KB)
• Preprint (1514 KB)
• Metadata XML

Status : open (until 02 Jul 2024)

Report abuse

Please provide a reason why you see this comment as being abusive. You might include your name and email but you can also stay anonymous.

Please provide a reason why you see this comment as being abusive.

Please confirm reCaptcha.

Sulav Ghimire

Gabriel m. g. guerreuro, kanakesh vatta kkuni, emerson david guest, kim høj jensen, guangya yang, xiongfei wang.