A Review on the Application of Artificial Intelligence in Anomaly Analysis Detection and Fault Location in Grid Indicator Calculation Data

Abstract

With the rapid development of artificial intelligence (AI), AI has been widely applied in anomaly analysis detection and fault location in power grid data and has made significant research progress. Through looking back on traditional methods and deep learning methods in anomaly analysis detection and fault location of power grid data, we aim to provide readers with a comprehensive understanding of the existing knowledge and research advancements in this field. Firstly, we introduce the importance of anomaly analysis detection and fault location in power grid data for the safety and stability of power system operations and review traditional methods for anomaly analysis detection and fault location in power grid data, analyzing their advantages and disadvantages. Next, the paper briefly introduces the concepts of commonly used deep learning models in this field and explores, in depth, the application of deep learning methods in anomaly analysis detection and fault location of power grid data, summarizes the current research progress, and highlights the advantages of deep learning over traditional methods. Finally, we summarize the current issues and challenges faced by deep learning in this field and provide an outlook on future research direction.

1. Introduction

With the rapid development of the economy, the demand of electricity is constantly rising, showing an increasing trend year by year [1]. This growth trend stems from the extensive application of electricity in modern life, covering various fields such as industry, commerce, and households. The acceleration of technological progress and urbanization has made electrical energy an indispensable infrastructure for the operation of modern society. Therefore, in order to adapt to societal development, the scale of the power grid system has rapidly expanded, as shown in Figure 1. This leads to an exponential increase in the volume of operational data at every stage of power grid operation. And this will lead to increasing complexities in grid operation data, making data anomalies and fault issues more prominent.

Data anomalies are a common quality issue in statistical data across various industries, and the presence of anomalous data in power grid data is a key factor affecting data quality [2,3]. To address this issue, researchers have employed various methods for anomaly detection in power grid data, such as the Isolation Forest algorithm and Density Peaks Clustering algorithm. However, due to the large scale and complexity of power grid data, the accuracy and efficiency of traditional anomaly detection methods are relatively low.

Since the advent of power networks, fault location in power systems has always been a significant concern for researchers. A fast and accurate fault location method can help to improve the continuity of the supply considerably [4]. Researchers have conducted studies on grid fault location issues using traditional methods such as matrix operations and models based on random forest regression. However, these traditional methods may have errors and uncertainties, affecting the accuracy and reliability of fault location.

Deep learning mainly refers to a class of algorithms that learn the characteristics of things from large amounts of data. By constructing multi-layer network models, it automatically extracts useful information from data to classify or predict data [5]. With the rapid development of deep learning, neural networks have demonstrated excellent performance in various fields, including image classification [6], natural language processing [6], speech recognition [7], and medical image processing [8]. Given the powerful advantages of neural networks in data feature extraction, researchers have introduced deep learning technology into the field of power grid anomaly analysis, detection, and fault location, expecting neural networks to fully extract the deep features of power grid data to achieve more accurate results.

Next, the first chapter of this paper introduces the application of traditional methods in power grid data anomaly analysis, detection, and fault location. The second chapter briefly reviews the basic concepts of deep learning models, focusing on the application of deep learning in power grid data anomaly analysis, detection, and fault location. The third chapter sorts out the current problems and challenges faced by power grid data anomaly detection and fault location. Finally, this paper summarizes and prospects the future development trends in this field.

2. Traditional Methods

The concepts of power grid anomaly analysis, detection, and fault location were proposed in the early stages of power system development. As power systems continued to develop and expand, the importance of anomaly analysis detection and fault location became increasingly recognized. However, the earliest methods lacked scientific approaches and technical support, resulting in low accuracy and efficiency. With the continuous advancement of science and technology, traditional analysis methods and techniques were gradually introduced into power grid data anomaly analysis, detection, and fault location.

2.1. The Application of Traditional Methods in Power Grid Data Anomaly Analysis and Detection

Rodrigues et al. [9] introduced a novel density-based clustering algorithm known as the Density Peaks Clustering (DPC) algorithm. This innovative approach involves plotting a decision graph based on the characteristics of cluster centers, which enables the algorithm to quickly and accurately identify these centers [10]. Building upon this, Lu et al. [11] applied the DPC algorithm to power data anomaly detection and demonstrated, through experimental results, that it significantly improved the detection efficiency. In addition to the DPC algorithm, other researchers have explored various clustering algorithms for power data anomaly detection. For instance, Wu et al. [12] utilized an enhanced K-means algorithm, Wang et al. [13] employed an improved Fuzzy C-Means (FCM) algorithm, and Li et al. [14] implemented an upgraded Particle Swarm Optimization-based Fuzzy C-Means (PSO-PFCM) clustering algorithm, each contributing to advancements in anomaly detection methods. To reduce the computation time in power data anomaly detection, Gao et al. [15] proposed an improved spectral clustering algorithm that prunes data to a lower complexity. Experiments with real power data showed that this method outperforms existing approaches, as evidenced by the results in

Table 1. Algorithm accuracy.

Algorithm	Accuracy (%)↑
Spectral clustering [15]	86.50
K-means [15]	87.55
Mini-batch K-means [15]	88.50
Improved algorithm [15]	92.50

. In this context, “↑” indicates that a higher value is better, while “↓” indicates that a lower value is better. Clustering algorithms significantly enhance power data anomaly detection by grouping data and identifying anomalies, thus improving the accuracy and efficiency. For instance, the DPC algorithm uses a decision graph for quick anomaly detection, while K-means, improved FCM, and PSO-PFCM optimize clustering results. Despite these advancements, choosing the right algorithm and parameters is crucial for effective detection.

The Isolation Forest algorithm is a randomized anomaly detection algorithm that leverages the isolation of normal samples relative to anomalous ones for detection. The basic idea of the algorithm is to construct a random binary tree to separate normal samples from anomalous samples. The formula for calculating the anomaly score in the Isolation Forest algorithm is:

$$s(x,n)={\displaystyle \frac{E\left(h\right(x\left)\right)}{c\left(\mathsf{\Psi }\right)}}$$

where the anomaly score s(x,n) is calculated by dividing the average path length E(h(x)) of the data point x across n trees by the average path length c(Ψ) of all samples. The smaller this ratio, the more likely the data point is an anomaly. Here, x represents the data point, n represents the number of trees in the Isolation Forest, Ψ represents the total number of samples, h(x) indicates the path length of the data point x, E(h(x)) is the average path length of x, and c(Ψ) is the average path length of all samples.

Researchers have extensively explored the Isolation Forest algorithm for anomaly detection. For instance, Niu et al. [16] developed a method to improve the accuracy in power data anomaly detection using Isolation Forest. Li et al. [17] proposed a scheduling flow data anomaly detection method based on this algorithm. Similarly, Luo et al. [18] introduced an attribute-associated Isolation Forest algorithm for detecting anomalies in electro-data, while Mao et al. [19] designed an unsupervised model for power data anomaly detection primarily using Isolation Forest. Additionally, Li et al. [20] combined the Isolation Forest algorithm with the Local Outlier Factor algorithm to enhance power data anomaly detection through data mining techniques. The Isolation Forest algorithm excels in power data anomaly detection by using random trees to measure path lengths and identify anomalies. It is efficient and has a low computational complexity, making it ideal for real-time detection. Combined with other algorithms, it enhances the accuracy, proving valuable for power data analysis.

Support vector machine (SVM) is a supervised learning algorithm primarily used for classification tasks but also applicable to regression problems. It excels in handling high-dimensional data by constructing an optimal hyperplane to effectively distinguish between normal and anomalous data. In power data anomaly analysis and detection, SVM has demonstrated strong capabilities. For example, Liu et al. [21] proposed an SVM-based method for detecting anomalies in power system scheduling data, accurately identifying abnormal patterns. Additionally, Yu et al. [22] improved SVM technology to enhance the detection accuracy and reliability while reducing false negatives. Experimental results (

Table 2. Comparison results.

Method	Traditional Multi-Domain Feature Extraction [22]	Traditional Clustering Algorithm [22]	Improved SVM [22]
Abnormal difference value↓	1.25	1.03	0.67
Location response time(s)↓	2.12	2.35	1.03
Dynamic precision ratio↓	4.30	5.20	2.10
Missing report rate (%)↓	7.15	10.31	2.04

) show that this improved method offers greater accuracy and reliability in complex power environments, highlighting its practical application value. Overall, SVM’s efficiency and accuracy make it a crucial tool for power data anomaly analysis and detection.

Cloud computing represents a significant advancement in IT, providing on-demand network access to scalable IT resources such as servers, storage, and applications [23]. In the context of power data anomaly analysis and detection, cloud computing has proven highly effective. Shi et al. [24] proposed a novel cloud computing-based rapid anomaly detection algorithm that enhances the detection efficiency while maintaining high accuracy. Lu et al. [25] developed a method for analyzing and detecting anomalies in power dispatch data using a cloud computing platform, with simulation experiments demonstrating improved accuracy and efficiency in anomaly detection. In summary, cloud computing not only optimizes the anomaly detection process but also significantly enhances the practical application value of power data analysis.

2.2. The Application of Traditional Methods in Power Grid Fault Location

Precise information on fault location plays a vital role in expediting the restoration process after being subjected to any kind of fault in the power grid [26]. Since the beginning of the 21st century, research on power grid fault location has achieved certain results. For example, Mu et al. [27] proposed a method for fault location in microgrids that identifies the fault point through matrix operations, thereby determining the fault location. El Mrabet et al. [28] proposed a data-driven approach for determining fault characteristics using samples of fault trajectories. A random forest regressor-based model is used to detect the real-time fault location and its duration simultaneously. The experimental results indicate that the proposed model (RFR) outperforms the seven other models in terms of fault localization accuracy at nine different fault locations. Guo et al. [29] proposed a regional power grid fault location method that can utilize device data from multiple stations within a regional power grid for complete fault location, thereby improving system reliability and reducing costs. Galvez et al. [30] presented a robust fault location method that can be used for radial or meshed power systems. This is accomplished via a K-means clustering of DFRs and using a weighted directed tree model, enabling accurate fault location.

Caporuscio et al. [31] proposed a data-driven ground fault location method for the power distribution system. According to the results, the method shows good potential, with a total relative error of 0.4% for fault distance prediction. Mohammadi et al. [32] proposed a wide-area fault location method based on the Weighted Maximum Exponential Square (WMES) algorithm, which is highly robust to measurement uncertainties. The results of a simulation on the standard IEEE 57-bus network illustrate the efficiency of the proposed fault location method based on the WMES. Hassani et al. [33] proposed a novel hybrid framework for fault detection, identification, and location in power systems. Furthermore, in order to make a reliable decision as well as mitigating false alarms generated in the detection module, a novel zGT2FFM has been proposed. Yang et al. [34] proposed a fault section location method for active distribution networks based on the wolf pack and differential evolution algorithms. They introduced the differential evolution algorithm to the wolf pack algorithm to enrich the population diversity and enhance the global optimization performance. Zou et al. [35] proposed a comprehensive fault location method based on weighted least squares. This method utilizes the magnitudes of three-phase voltages at fault points and the impedance of fault line segments between the fault point and both ends of the fault line to expand the unknown variable vector.

The methods used in these references, though varied, share several notable advantages: they are mature and stable, simple to operate, and easy to understand, with broad applicability. They are cost effective and offer strong real-time performance, aiding in the rapid restoration of power grid operations. Additionally, they are relatively easy to maintain and operate, requiring minimal technical support, which makes them valuable in practical applications.

2.3. The Limitations of Traditional Methods

The traditional methods for power grid data anomaly analysis, detection, and fault location have exhibited a certain level of maturity and stability compared to earlier manual detection methods. However, with the increasing complexity of power grids and the expansion of the data scale, the limitations of traditional methods have gradually become apparent, including the following:

• Difficulty in handling massive data:

As the scale of power grids expands and the adoption of smart meters and sensors increases, the amount of grid data grows exponentially. Traditional methods often face limitations in their computational efficiency and storage capacity when processing and analyzing massive data, making real-time anomaly detection and fault localization challenging.

• Weak nonlinear relationship handling:

The variables in grid data often exhibit complex nonlinear relationships. Traditional methods like linear regression perform poorly in capturing and modeling these nonlinear relationships.

• Lack of self-learning capability

Traditional methods lack self-learning capabilities, unable to automatically learn from new data and update models. This means that when the grid environment or data patterns change, traditional methods require manual intervention for adjustments, resulting in slow response times.

These limitations may include issues related to accuracy, efficiency, and adaptability in anomaly analysis, detection, and fault location. Therefore, integrating advanced technologies such as deep learning can further enhance the performance of power grid data anomaly analysis, detection, and fault location.

3. Deep Learning Methods

With the rapid development of deep learning, its application in the analysis, detection, and localization of data anomalies in power grids has gradually become the focus of research and practice, bringing new possibilities and innovative solutions to this field. The limitations of traditional methods have been discussed earlier, while deep learning, as an emerging technology, offers many advantages compared to traditional methods, as shown in

Table 3. Comparison between traditional methods and deep learning.

Method	Advantages	Disadvantages
Traditional methods	Strong interpretability, relatively low computational cost	Difficult to handle large-scale data
Deep learning methods	Strong learning ability, capable of handling large-scale data	High data and computational resource requirements with lower interpretability

in a comparison of advantages and disadvantages.

3.1. The Application of Deep Learning in Power Grid Data Anomaly Analysis and Detection

3.1.1. Convolutional Neural Network

A convolutional neural network (CNN) is one of the most significant networks in the deep learning field [36]. There are usually several convolutional and pooling layers, which are arranged alternately; that is, a convolutional layer is connected to a pooling layer, followed by another convolutional layer after the pooling layer, and so on [37]. In the field of power grid data anomaly analysis and detection, CNN models can also be utilized for anomaly detection. Danilczyk et al. [38] implemented a CNN to analyze bus voltage data in a distributed power system. This anomaly detection architecture was able to correctly detect faults in the distributed power system that lasted only fractions of a second. Additionally, the CNN was able to accurately locate the power fault within the system.

Huang et al. [39] proposed a self-supervised, learning-based electricity abnormality detection model based on time convolutional networks. The experimental results, as shown in

Table 4. Comparison and analysis of test results.

Algorithm	SVM [39]	CNN [39]	RNN [39]	Ours-Sup [39]
Recall↑	0.633	0.838	0.836	0.908
AUC↑	0.633	0.803	0.811	0.889
F1-Score↑	0.663	0.781	0.829	0.901
Accuracy (%)↑	69.80	73.30	82.40	89.60

, demonstrate that the performance of this algorithm surpasses those of SVM, Recurrent Neural Network (RNN), and CNN. The evaluation metrics used are Area Under the Curve (AUC), accuracy, recall, and F1-Score. AUC evaluates binary classification models by measuring the area under the Receiver Operating Characteristic curve, with values from 0 to 1. AUC values closer to 1 indicate better performance. Accuracy is the rate of correct anomaly detection by the model. Recall, also known as sensitivity, measures the proportion of actual positive cases that are correctly identified by a model. F1-Score is the harmonic mean of precision and recall, used to evaluate the overall performance of a classification model.

3.1.2. Deep Neural Network

Deep neural network (DNN) has gained unprecedented performance due to its automated feature extraction capability [40]. This high-level performance has led to the application of DNN models in power grid data anomaly analysis and detection. Chang et al. [41], aiming to achieve a higher data recognition efficiency, proposed an electricity data anomaly detection technique based on improved K-means and DNN algorithms combined with Spark architecture theory. The simulation results demonstrate that their method efficiently handles anomalies in large-scale power grid operational data. A Generative Adversarial Network (GAN) consists of two neural networks, a generator and a discriminator, trained together to produce realistic data. Siniosoglou et al. [42] combined it with DNN to propose MENSA (anoMaly dEtection aNd claSsificAtion), a model capable of detecting 13 types of Modbus/TCP cyberattacks, 5 types of DNP3 cyberattacks, and potential anomalies related to operational data (i.e., time-series electricity measurements).

3.1.3. Long Short-Term Memory Network

Hochreiter et al. [43] proposed the Long Short-Term Memory (LSTM) network in 1997, with a key innovation being the introduction of three gates (input gate, forget gate, and output gate) and a cell state, enabling the network to effectively capture and memorize long-term dependencies in sequences. Zhang et al. [44] utilized LSTM autoencoders to perform anomaly analysis and detection on power grid data. Wu et al. [45], based on the inherent characteristics of data and LSTM theory, proposed a data-driven abnormal detection algorithm for user electricity data, and the experimental results showed an improved performance compared to traditional methods. Given the temporal characteristics present in most power grid datasets, Guha et al. [46] explored an LSTM-variational autoencoder-based deep generative model that can handle the moderate presence of anomalous data during training instead of standard data. This research highlights the advantages of reconstruction-based methods over clustering-based methods.

To enhance the performance of smart grids, Parsai et al. [47] proposed an innovative anomaly detection and power consumption prediction method using LSTM. When compared to the Autoregressive Integrated Moving Average algorithm (ARIMA), this method achieved a 20% reduction in the forecasting error. To address the data imbalance problem caused by the lack of anomalous data, Zhou et al. [48] proposed a data-driven framework based on a LSTM to directly detect anomalies in distribution systems using voltage magnitude measurements. This method successfully detected all predefined anomalies with a high degree of confidence.

3.1.4. Recurrent Neural Network

RNNs are designed for sequential data processing. They use recurrent connections to transmit information between time steps, enabling the memory of previous inputs. Building on this foundation, Scholar et al. [49] proposed an intelligent deep learning method for detecting anomalies in power grids. This approach employs the Modified Flow Direction Algorithm (MFDA) for optimal feature selection, which are then processed using an Adaptive Residual RNN with a Dilated, Gated Recurrent Unit (ARRNN-DGRU) for identification. The simulation results highlight the model’s superior performance, showing improved detection rates and the enhanced robustness of the smart grid system compared to existing methods. To improve poor anomaly detection performance due to the spatio-temporal and multi-dimensional nature of measurement data, Zheng et al. [50] proposed an anomaly detection model based on an encoder–decoder framework and RNN, as shown in Figure 2. The model detects anomalies via high reconstruction error. The experimental results show that it outperforms state-of-the-art models, achieving over 95% precision.

3.2. The Application of Deep Learning in Power Grid Fault Localization

3.2.1. Convolutional Neural Network

A CNN is also applied in the field of power grid fault location, such as the framework proposed by Patel et al. [51] for locating line fault sections, as shown in Figure 3. This framework is designed to handle the detection, classification, and location of the fault without using the system parameters. The framework proposed utilizes only the current data from the DFR at the receiving end of the line. First, Isolation Forest is used to detect the fault, and then, SVM is used to classify the fault. After this, a CNN is used to locate the fault. Given the weaknesses of fault features in the distribution network, Shi et al. [52] proposed a fault location method based on a one-dimensional convolutional neural network (1D-CNN). The model takes the timing signal of the electrical quantity measured at the end of the line as input and the estimated value of the fault point as output, which can mine the complex mapping relationship between fault features in the collected electrical quantity signals and the distance to the fault point, resulting in excellent performance and an acceptable fault location error.

Inspired by the Fourier transform, Yu et al. [53] proposed SIG-CNN, an online data-driven method that transforms signals from the time domain to the image domain using the Signal-to-Image (SIG) algorithm and then processes the images with a CNN framework. Under ideal conditions, comparisons with a BPANN, LSTM, and traditional CNN demonstrate the robustness and reliability of a SIG-CNN. As shown in

Table 5. Comparison of different methods.

Method	Traditional CNN [53]	BPNN [53]	LSTM [53]	SIG-CNN [53]
MAPE↓	0.880%	1.750%	1.120%	0.780%
RMSE↓	0.2371	0.3601	0.2833	0.1725
Accuracy↑	99.91%	98.74%	99.75%	99.99%

, the evaluation metrics used include accuracy, as well as mean absolute percentage error (MAPE) and root mean square error (RMSE) for fault regression problems [54].

3.2.2. Deep Neural Network

Given the excellent performance of DNNs, researchers have introduced them into the field of power grid fault location and conducted related studies. For example, Sapountzoglou et al. [55] proposed a DNN-based method for fault detection and localization in low-voltage smart distribution networks. This method addresses limitations in existing approaches by being grid topology independent, branch independent, effective with limited data, and uniquely accurate in detecting high-impedance faults. The experimental results show the method detects faulty feeders with a 100% accuracy, identifies faulty branches with an average accuracy of 84%, and estimates fault locations within branches with an average error of 12%. Luo et al. [56] proposed a fault location method for DC distribution networks based on deep learning. Initially, a DC distribution network with a radial topology is modeled, and faults with various parameters are introduced to simulate different scenarios encountered in practical projects. Subsequently, a deep neural network is constructed and trained using normalized fault currents. The parameters of the network are adjusted based on specific application requirements. Finally, the fault localization performance of the deep neural network is evaluated.

3.2.3. Artificial Neural Network

Artificial neural networks (ANNs) are a network structure that can be used to address real-world problems with multiple nodes and multiple output points. They gained recognition after AlphaGo defeated the world Go champion Lee Sedol in 2016, demonstrating the immense potential of artificial neural networks [57]. Based on this, Dashtdar et al. [58] proposed a method for fault location in distribution networks using an ANN. The study utilized two neural networks for fault identification and location, generating training data through wavelet transform of relay-observed fault signals and calculating the entropy of three-phase currents along with positive and zero-sequence energies. The comparison between the predicted values and the actual values indicates that the proposed method has a high accuracy in locating fault regions. Barra et al. [59] proposed a new method based on ANNs to locate faulted areas in radial distribution networks using voltage and current measurements at the substation. Multiple simulations conducted using PSCAD™/EMTDC™ 5.0 software demonstrated that the proposed method has a high level of accuracy in locating faulted areas. Meanwhile, Usman et al. [60] classified different types of faults while locating various faults in the distribution network using an ANN. The proposed method can classify all types of faults that may occur in the grid and then identify the approximate fault location based on the fault type. Testing was conducted on a modified IEEE-37 bus test case, and the results showed a 100% accuracy in classifying all types of faults and an over 99% accuracy in identifying the fault location.

The Backpropagation Neural Network (BPNN) is a widely used ANN mode. It consists of an input layer, hidden layers, and an output layer and utilizes the backpropagation algorithm to continuously adjust connection weights based on training data, thereby modeling and predicting complex nonlinear relationships. Han et al. [61] applied both the traditional BPNN and optimized BP to the design of distribution network systems and simulated them using MATLAB R2014a software. The experimental results showed that the performance of the BPNN improved after optimization using genetic algorithms compared to the traditional BPNN. Li et al. [62] integrated the BPNN with cloud genetic algorithms to study fault localization in distribution networks. The experimental results demonstrated that the optimized BPNN using cloud genetic algorithms could be effectively utilized for fault localization in distribution networks, thereby enhancing the accuracy and effectiveness of fault localization.

3.2.4. Graph Convolutional Network

Graph Convolutional Networks (GCNs) are efficient models for processing graph data and have attracted widespread attention from researchers in recent years [63]. Chen et al. [64] presented a novel Graph Convolutional Network framework for fault location in power distribution networks. The proposed method integrates multiple measurements from different buses while considering the system topology. The simulation results, as shown in

Table 6. Fault location accuracies of different approaches.

Model	PCA + SVM [64]	PCA + RF [64]	FCNN [64]	GCN [64]
Accuracy(%)↑	94.60	94.60	84.64	99.26
One-hop accuracy(%)↑	98.31	99.28	96.38	99.93

, indicate that the GCN model significantly outperforms other widely used machine learning methods, achieving a very high fault localization accuracy. The experimental comparison methods include principal component analysis (PCA), support vector machine (SVM), random forest (RF), and a fully connected neural network (FCNN).

4. Advantages and Limitations

4.1. Advantages

AI has several advantages in anomaly analysis, detection, and fault location in power grid data, including the following:

4.1.1. Automated Feature Extraction

AI can automatically identify and extract important features from power grid data, thereby simplifying the analysis process. Utilizing advanced neural network models, AI is able to learn and recognize complex feature representations without the need for manual selection or intervention. This innovative approach enables the model to discover subtle patterns that traditional methods often overlook, including anomalies in equipment operation and the impact of environmental changes on grid performance. By automating feature extraction, it significantly reduces the labor costs and processing time, thereby enhancing the overall efficiency and effectiveness of data analysis.

4.1.2. Efficient Handling of Massive Data

As the scale of power grids continues to expand, the volume of data generated grows exponentially, presenting significant challenges for traditional data-processing methods. However, AI’s distributed computing capabilities provide a powerful solution, allowing it to efficiently manage and analyze massive amounts of data. By leveraging advanced platforms such as cloud computing, AI can process data from a diverse array of sensors and devices in parallel, facilitating real-time analysis. This capability not only enhances the speed and efficiency of data processing but also ensures that the grid system remains operational under high load conditions. Consequently, it can monitor the grid status in real time, promptly identifying potential anomalies and faults, thereby improving the overall reliability and safety in power distribution.

4.1.3. Self-Learning and Adaptability

AI systems possess remarkable adaptive capabilities, enabling them to automatically adjust to new data and changes in their operating environment. Through continuous learning processes, these models can progressively enhance their detection performance over time, becoming more accurate and efficient. For example, when there are shifts in the grid equipment or variations in the operational patterns, AI can seamlessly update its internal parameters to maintain optimal performance. This self-learning ability not only allows AI to keep pace with the constantly evolving power environment but also reduces the need for frequent manual intervention, ultimately enhancing the reliability and responsiveness of the system in real-world applications.

4.2. Limitations

4.2.1. Data Quality Dependency

In power grid data analysis, the performance of artificial intelligence models heavily relies on the quality of input data. Although power grid systems generate large volumes of data, these data are often affected by noise, missing values, and erroneous labels. For example, sensors may record inaccurate data due to environmental interference, leading the model to mistakenly identify normal states as anomalies, which can trigger unnecessary alarms. Therefore, ensuring the accuracy, completeness, and timely updating of power grid data is crucial. Only with high-quality data can AI technology genuinely enhance the safety and reliability of the power grid, ensuring its efficient operation.

4.2.2. Model Generalization Ability

The operating environment and conditions of power grids are complex and continually evolving, with significant variations across different regions and seasons. Models typically rely on specific historical data during the training process. If a model cannot generalize effectively to new operating conditions and environments, its ability to detect and localize faults will be limited. Furthermore, there may be new types of anomalies and faults that have not occurred or are rarely present in historical data, complicating the model’s capacity to identify and manage these emerging issues effectively.

4.2.3. Lack of Interpretability

Many AI algorithms, particularly deep learning models, face a significant lack of interpretability. This means that when the model detects anomalies and faults, it cannot clearly articulate the reasons behind its decisions or provide a transparent rationale. In power system applications, this lack of clarity is especially concerning, as operators need to understand the decision-making process of the model in order to respond quickly and effectively in emergency situations. Without interpretability, decision makers may struggle to trust the model’s predictions, which ultimately hinders the practical application and overall effectiveness of AI solutions in grid infrastructure.

5. Future Research Directions

5.1. Advanced Models

With the rapid development in the fields of deep learning and machine learning, it is expected that the new generation of models will fully leverage more complex structures such as convolutional neural networks or attention mechanisms to capture the characteristics of power grid data more accurately and finely. The use of complex models is anticipated to enhance the understanding of abnormal and fault data, providing more precise tools for power grid data anomaly analysis detection and fault localization. However, balancing the complexity of the models with computational efficiency and addressing the challenges of handling large-scale data will be important issues that future research needs to address.

5.2. Transfer Learning and Few-Shot Learning

Implementing transfer learning and few-shot learning techniques can enable models to quickly adapt to new power grid data. Transfer learning [65,66] aims to improve the learning performance on the target domain via reusing the knowledge learned from source domains. Through transfer learning, models can leverage knowledge and features learned in one domain to accelerate the learning process in another, thereby reducing the reliance on large amounts of labeled data and improving the model’s generalization and adaptability. Few-shot learning focuses on situations where labeled data are scarce, effectively utilizing limited data for model training and optimization, thus maximizing the value of the data and achieving rapid adaptation and effective utilization of new datasets. The application of these technologies helps to enhance the model’s generalization ability in the power grid, providing stronger support for anomaly analysis detection and fault localization in power grid data.

5.3. Federated Learning

As machine learning technology continues to advance, federated learning, as an emerging distributed machine learning method, is expected to play a greater role in the detection of anomalies and fault location in power grid data. One of the core advantages of federated learning is its ability to perform data analysis while protecting data privacy. With the strengthening of data security regulations, federated learning will provide power grid companies with a secure platform for data sharing and analysis, enabling the detection and location of power grid anomalies without leaking user data. In the future, federated learning may be combined with other field technologies, such as the Internet of Things (IoT) and big data analysis, to form a more comprehensive smart grid management system. This will help achieve more comprehensive monitoring and faster responses.

5.4. Multi-Source Data Integration

The future prospects of multi-source data integration in anomaly analysis, detection, and fault location in power grids are promising. With the development of smart sensors, smart meters, and IoT technologies, the grid will be able to collect real-time data from various sources, including equipment status, weather information, user behavior, and historical fault records. This comprehensive data integration will enable more accurate anomaly analysis and detection, leveraging deep learning algorithms to analyze multidimensional data for rapid fault localization. Additionally, with the self-learning capabilities of artificial intelligence, the system can continually optimize its models, proactively identify risks, and recommend maintenance actions. The integrated data will also support real-time decision making, helping grid operators respond quickly to anomalies and reduce fault recovery time. Furthermore, by incorporating user data, power grid management will become more intelligent, enhancing the resource utilization efficiency and achieving sustainable development.

6. Conclusions

In recent years, the national requirements for the safety and stability of power grids have been increasing, making anomaly analysis, detection, and fault location particularly important. The timely detection of anomalies and prevention of faults have become key research focuses. The research on anomaly detection and fault localization in power grid data is continuously evolving. Traditional methods often require substantial manpower and have certain limitations. In the context of rapid technological development, the rise of AI has introduced new possibilities for anomaly analysis, detection, and fault location in power grids, significantly improving the accuracy and efficiency through AI methods. We comprehensively discuss anomaly detection and fault localization in power grid data by integrating traditional methods and deep learning approaches. By deeply analyzing the advantages and challenges of AI in this field and providing a profound outlook on future research directions, the paper aims to give readers a more comprehensive understanding of the topic. Additionally, this paper seeks to provide valuable references and insights for future research, encouraging more breakthroughs in this field.