Power distribution system fault analysis method and system

Abstract

The invention discloses a power distribution system fault analysis method and system. The method comprises the steps of collecting operation related data of a power distribution system in real time through various sensors and intelligent metering equipment; preprocessing the collected operation related data to remove noise and interference signals, and establishing a multi-dimensional feature set of the real-time operation state of the power distribution system; extracting potential fault features from the multi-dimensional feature set by using a convolutional neural network, and generating a fault feature vector; performing time sequence analysis on the fault feature vector by using a recurrent neural network, and identifying a possible fault type in the current power distribution system; and in combination with the topological structure and the electrical model of the power distribution system, spatial positioning is performed on the identified fault type based on the graph neural network, and specific nodes of the fault and corresponding fault severity levels are output. According to the embodiment of the invention, the fault detection capability and response speed of the power distribution system can be remarkably improved, and the power distribution system is promoted to develop towards intelligentization and automation directions.

Description

Technical Field

[0001] The present invention belongs to the technical field of power distribution, and in particular to a method and system for analyzing power distribution system faults. Background Art

[0002] With modern society's increasing reliance on electricity supply, the reliability and safety of distribution systems have become a critical research topic in the power industry. As a key link in power transmission, distribution systems are responsible for delivering electricity from substations to users. However, over the long term, distribution systems can be susceptible to a variety of factors, leading to failures such as sudden load changes, equipment aging, and environmental changes. If these failures are not promptly detected and addressed, they can cause power outages, damage equipment, and even pose safety risks, resulting in significant socioeconomic losses.

[0003] Traditional distribution system fault monitoring and location rely primarily on manual inspections and basic monitoring equipment. This approach is not only inefficient but also difficult to accurately diagnose complex faults. With the development of intelligent technology, real-time data collection through various sensors and smart metering devices has become an effective means of improving fault detection capabilities. However, relying solely on direct analysis of monitoring data often makes it difficult to fully understand the operating status and potential faults of the distribution system, resulting in insufficient accuracy and timeliness in fault detection. Summary of the Invention

[0004] The purpose of the present invention is to provide a distribution system fault analysis method and system to address the deficiencies in the prior art, significantly improve the fault detection capability and response speed of the distribution system, and promote the development of the distribution system towards intelligence and automation.

[0005] An embodiment of the present application provides a method for analyzing a power distribution system fault, the method comprising:

[0006] Collect distribution system operation-related data in real time through various sensors and smart metering devices;

[0007] Preprocessing the collected operation-related data to remove noise and interference signals and establish a multi-dimensional feature set of the real-time operation status of the power distribution system, wherein the multi-dimensional feature set includes electrical characteristics, environmental characteristics, and load characteristics;

[0008] A convolutional neural network is used to extract potential fault features from the multidimensional feature set and generate a fault feature vector, wherein the fault features include at least: load mutation, harmonic interference and overvoltage;

[0009] Using a recurrent neural network to perform time series analysis on the fault feature vectors to identify possible fault types in the current power distribution system, where the fault types include short circuit faults, ground faults, and equipment aging faults;

[0010] Combining the topological structure and electrical model of the power distribution system, the identified fault type is spatially located based on the graph neural network, and the specific node where the fault occurs and the corresponding fault severity level are output.

[0011] Optionally, the collected operation-related data is preprocessed to remove noise and interference signals and establish a multi-dimensional feature set of the real-time operation status of the power distribution system, including:

[0012] The current and voltage signals in the collected operating data are decomposed using wavelet transform to extract low-frequency and high-frequency components, and high-frequency noise is removed through threshold processing to retain the main features of the signal;

[0013] For missing values that may appear during the data collection process, an interpolation method based on the K-nearest neighbor algorithm is used to fill them according to the characteristic values of similar time points to ensure the integrity of the operating data;

[0014] Principal component analysis is used to reduce the dimensionality of the interpolated operating data, extracting the main electrical characteristics, environmental characteristics, and load characteristics to form a preliminary feature set. This preliminary feature set is then enhanced using a generative adversarial network to generate diverse feature samples.

[0015] Using a random forest algorithm to evaluate the contribution of each feature in the feature sample to fault prediction, and generating a feature importance score as a feature weight;

[0016] The feature weights are used to weight each feature, and the weighted features are fused through a multi-layer perceptron to form a comprehensive multi-dimensional feature set. The multi-dimensional feature set can capture the real-time operating status of the power distribution system and reflect the relationship between electrical characteristics, environmental characteristics and load characteristics.

[0017] Optionally, the method uses a convolutional neural network to extract potential fault features from the multidimensional feature set and generate a fault feature vector, wherein the fault features include at least load mutation, harmonic interference, and overvoltage, including:

[0018] Construct a hierarchical convolutional neural network whose structure consists of multiple convolutional layers, pooling layers and fully connected layers; the network is divided into three main modules:

[0019] Local feature extraction module: Through multiple convolutional layers, small convolution kernels are used to extract local features from the multidimensional feature set to obtain local feature maps to identify basic signal waveform changes. Signal waveform changes include instantaneous mutations and waveform distortions in current and voltage. The convolution operation in this layer is designed to capture subtle changes in the signal.

[0020] After each convolution layer is convolved, a residual connection is used to add the local feature map output by the convolution layer to the features in the multidimensional feature set input to the convolution layer at the element level to obtain a residual connection feature map;

[0021] Feature Abstraction Module: Continuing to use convolutional layers and max pooling layers, the network gradually reduces the dimensionality of the residual connection feature map, enabling it to abstract high-level features from local features to obtain a high-level feature map. These high-level features include harmonic components and periodic fluctuation patterns. This module uses larger convolution kernels to help the model identify trends and periodic features in the signal.

[0022] High-level feature synthesis module: Uses a global average pooling layer to convert the spatial dimensions of all high-level feature maps into one-dimensional vectors to retain the important information of each high-level feature map and form a comprehensive feature vector;

[0023] The comprehensive feature vector is passed to the fully connected layer to further learn the relationship between features. The output vector generated by the softmax activation function is used as the fault feature vector. Each dimension of the fault feature vector corresponds to a different fault feature, where the fault features include at least: load mutation, harmonic interference and overvoltage.

[0024] Optionally, the method of using a recurrent neural network to perform time series analysis on the fault feature vector to identify possible fault types in the current power distribution system includes:

[0025] Arrange the fault feature vectors in chronological order to form a time series data set;

[0026] A pre-trained multi-layer long short-term memory network is used to perform time series analysis on the time series data set, wherein the input is the fault feature vector at the current moment and the fault feature vectors of the previous few time steps in the time series data set, and the output is a probability distribution representing the possible fault types and their probabilities in the current power distribution system. In addition, the structure of the multi-layer long short-term memory network includes: an input layer, multiple LSTM layers and a fully connected layer. The input layer receives time series data, the LSTM layer is responsible for extracting time series features, and the final fully connected layer is used to output the probability distribution of the fault type.

[0027] Optionally, the method combines the topology and electrical model of the power distribution system to spatially locate the identified fault type based on a graph neural network, and outputs the specific node where the fault occurred and the corresponding fault severity level, including:

[0028] Construct a topology diagram of the power distribution system, where the nodes of the topology diagram represent the devices in the power distribution system, and the edges represent the electrical connections between the devices. The steps of constructing the topology diagram include:

[0029] Assign a unique node identifier to each device and record its electrical and environmental characteristics;

[0030] Define edge weights based on the electrical connection relationships of the power distribution system. The weights are determined based on the electrical characteristics of the devices and reflect the electrical impact between the devices.

[0031] Representing a topological graph as an adjacency matrix and a characteristic matrix, wherein the adjacency matrix represents the connection relationship between nodes and the characteristic matrix contains the electrical and environmental characteristics of each node;

[0032] The identified fault type is mapped to the nodes of the topology graph, and the mapped topology graph is input into the pre-trained graph neural network to output the specific node where the fault occurred and the corresponding fault severity level.

[0033] Another embodiment of the present application provides a power distribution system fault analysis system, the system comprising:

[0034] The acquisition module is used to collect real-time operation-related data of the power distribution system through various sensors and intelligent metering devices;

[0035] a processing module, configured to pre-process the collected operation-related data to remove noise and interference signals and establish a multi-dimensional feature set of the real-time operation status of the power distribution system, wherein the multi-dimensional feature set includes electrical characteristics, environmental characteristics, and load characteristics;

[0036] An extraction module, configured to extract potential fault features from the multidimensional feature set using a convolutional neural network and generate a fault feature vector, wherein the fault features include at least load mutation, harmonic interference, and overvoltage;

[0037] An identification module is used to perform time series analysis on the fault feature vector using a recurrent neural network to identify possible fault types in the current power distribution system, wherein the fault types include short circuit faults, ground faults, and equipment aging faults;

[0038] The output module is used to combine the topology and electrical model of the distribution system, spatially locate the identified fault type based on the graph neural network, and output the specific node where the fault occurred and the corresponding fault severity level.

[0039] Yet another embodiment of the present application provides a storage medium, wherein the storage medium stores a computer program, wherein the computer program is configured to execute any of the above methods when run.

[0040] Yet another embodiment of the present application provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute any of the above methods.

[0041] Compared with the existing technology, the present invention provides a distribution system fault analysis method, which collects operation-related data of the distribution system in real time through multiple sensors and intelligent metering equipment; preprocesses the collected operation-related data to remove noise and interference signals, and establishes a multidimensional feature set of the real-time operation status of the distribution system; uses a convolutional neural network to extract potential fault features from the multidimensional feature set and generates a fault feature vector; uses a recurrent neural network to perform time series analysis on the fault feature vector to identify the possible fault types in the current distribution system; combines the topological structure and electrical model of the distribution system, and spatially locates the identified fault type based on the graph neural network, and outputs the specific node where the fault occurs and the corresponding fault severity level, thereby significantly improving the fault detection capability and response speed of the distribution system, and promoting the development of the distribution system towards intelligence and automation. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 A hardware structure block diagram of a computer terminal for a power distribution system fault analysis method provided by an embodiment of the present invention;

[0043] Figure 2 A schematic diagram of a flow chart of a power distribution system fault analysis method provided by an embodiment of the present invention;

[0044] Figure 3 A schematic structural diagram of a power distribution system fault analysis system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0045] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and are not to be construed as limiting the present invention.

[0046] The embodiment of the present invention first provides a method for analyzing faults in a power distribution system. The method can be applied to electronic equipment, such as a computer terminal, specifically a common computer.

[0047] The following describes it in detail by taking running on a computer terminal as an example. Figure 1 The hardware structure block diagram of a computer terminal for a power distribution system fault analysis method provided by an embodiment of the present invention. Figure 1As shown, the computer device includes a processor, a memory, and a network interface connected via a system bus, wherein the memory may include a non-volatile storage medium and an internal memory.

[0048] The non-volatile storage medium can store an operating system and a computer program. The computer program includes program instructions, and when the program instructions are executed, the processor can execute any one of the power distribution system fault analysis methods.

[0049] The processor is used to provide computing and control capabilities and support the operation of the entire computer equipment.

[0050] The internal memory provides an environment for the operation of the computer program in the non-volatile storage medium. When the computer program is executed by the processor, the processor can execute any distribution system fault analysis method.

[0051] The network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art will understand that Figure 1 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0052] It should be understood that the processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0053] See also Figure 2 , an embodiment of the present invention provides a method for analyzing a power distribution system fault, which may include the following steps:

[0054] S201, collects real-time operation data of the power distribution system through various sensors and smart metering equipment;

[0055] In power distribution system fault analysis, real-time collection of system operational data through a variety of sensors and smart metering devices is a crucial step. This process involves using various types of sensors, such as current sensors, voltage sensors, temperature sensors, and environmental monitoring devices, to obtain the real-time operating status of the distribution system. These sensors can monitor various parameters, such as current, voltage, power factor, temperature, and humidity, ensuring comprehensive monitoring of the distribution system. For example, current sensors can record current changes in real time, while temperature sensors can monitor equipment operating temperature to promptly identify potential overheating issues. Through real-time data collection, the system can obtain accurate operating status information, providing a foundation for subsequent fault analysis and resolution.

[0056] Real-time data collection related to the operation of power distribution systems plays a crucial role. First, it provides essential data support for fault detection and analysis, enabling the system to promptly identify anomalies and reduce the risk of failure. Second, real-time data collection helps operations and maintenance personnel better understand the operating status of the power distribution system, optimize equipment maintenance and management strategies, and improve the overall reliability and safety of the system. Furthermore, analysis of collected data can identify potential failure modes and trends, providing a basis for intelligent system upgrades and driving the development of more efficient and intelligent power distribution systems.

[0057] This can be accomplished by deploying a comprehensive monitoring system. First, multiple sensors are installed at key nodes and equipment in the power distribution system. For example, current and voltage sensors are installed on transformers, distribution cabinets, and critical loads to monitor current and voltage changes in real time. Environmental sensors are also installed around the equipment to monitor environmental factors such as temperature and humidity. All sensors are connected to a central data acquisition system via a wireless or wired network to ensure real-time data transmission.

[0058] During data collection, the system collects sensor data at regular intervals (e.g., every second or every minute) and stores it in a database. To ensure data accuracy and integrity, the system also implements data validation and anomaly detection mechanisms to automatically identify and address possible sensor failures or data loss. For example, if a sensor fails to function properly, the system can interpolate data from adjacent sensors to ensure data continuity.

[0059] Furthermore, the data acquisition system can be integrated with cloud computing platforms, leveraging the cloud's powerful computing capabilities for data analysis and storage. Through data analysis, the system can generate real-time reports on the distribution system's operating status, helping operations and maintenance personnel quickly understand the system's health and take timely measures to address potential failure risks. Ultimately, this multi-sensor real-time data acquisition approach not only improves the distribution system's monitoring capabilities but also provides a solid data foundation for subsequent fault analysis and decision-making.

[0060] S202, preprocessing the collected operation-related data to remove noise and interference signals, and establishing a multi-dimensional feature set of the real-time operation status of the power distribution system, wherein the multi-dimensional feature set includes electrical characteristics, environmental characteristics, and load characteristics;

[0061] Preprocessing collected operational data is a critical step in ensuring accurate data analysis. In power distribution systems, signals from various sensors and metering devices may be subject to interference, including electrical noise, environmental changes, and errors in the sensors themselves. Left unprocessed, these noise and interference signals will affect subsequent feature extraction and fault identification. Therefore, this step employs a range of signal processing techniques to remove unnecessary interference and extract the most representative electrical, environmental, and load characteristics. For example, using wavelet transforms to process current and voltage signals effectively separates low-frequency signals from high-frequency noise, preserving the primary characteristics of the electrical signals. Furthermore, missing data points are filled using an interpolation method based on the K-nearest neighbor algorithm to ensure data integrity, thereby constructing a comprehensive multidimensional feature set.

[0062] The importance of implementing data preprocessing lies in improving the accuracy and reliability of distribution system fault analysis. First, by removing noise and interference, the resulting multidimensional feature set more accurately reflects the operating status of the distribution system, effectively avoiding misjudgments and omissions caused by data quality issues. Second, effective data preprocessing can significantly improve the training of subsequent deep learning models, enabling them to more accurately learn and identify potential fault characteristics. This helps operations and maintenance personnel identify potential fault risks early, allowing them to take appropriate preventive measures and improve the safety and stability of the distribution system.

[0063] Specifically, wavelet transform can be used to decompose the current and voltage signals in the collected operating data, extract low-frequency and high-frequency components, and remove high-frequency noise through threshold processing to retain the main features of the signal;

[0064] Wavelet transform is an effective time-frequency analysis tool that decomposes signals into frequency components, facilitating the removal of high-frequency noise. In power distribution systems, current and voltage signals are often affected by external electromagnetic interference or inherent noise from the equipment, generating high-frequency noise. By applying wavelet transform to the collected data, the system decomposes the signal into multiple frequency bands, extracting low-frequency and high-frequency components. Threshold processing techniques are then used to set a reasonable threshold to remove high-frequency noise while retaining low-frequency components, ensuring that the signal's key characteristics are not affected.

[0065] Using wavelet transform processing significantly improves signal quality and reliability, ensuring the accuracy of subsequent analysis. By removing high-frequency noise, the current and voltage signals acquired by the system more accurately reflect the operating status of the power distribution system. This clear signal helps maintenance personnel quickly identify equipment anomalies and take appropriate maintenance measures, thereby avoiding potential economic losses and safety hazards caused by failures.

[0066] In the specific implementation, first select a suitable wavelet function, such as the Daubechies wavelet, and apply it to the current and voltage signal data collected in real time. By selecting the appropriate decomposition level, usually 4 to 6 layers, the system decomposes the signal into multiple frequency components. During this process, an interface is provided for operation and maintenance personnel to set and adjust the threshold to ensure that the noise components removed are reasonable. Subsequently, a threshold denoising algorithm is used, including hard threshold and soft threshold processing, to remove high-frequency components above the set threshold, and the cleaned signal is reconstructed through an inverse wavelet transform. Ultimately, the current and voltage signals output by the system will only contain the necessary low-frequency components, which will serve as the basic data for subsequent feature extraction.

[0067] For missing values that may appear during the data collection process, an interpolation method based on the K-nearest neighbor algorithm is used to fill them according to the characteristic values of similar time points to ensure the integrity of the operating data;

[0068] During the data collection process, some data points may be missing due to various factors (such as sensor failures and network delays). To ensure the integrity of distribution system monitoring data, these missing values need to be effectively interpolated. Interpolation based on the K-nearest neighbor (KNN) algorithm is a common and efficient filling strategy. This method detects time points similar to the missing values and fills the missing data using a weighted average or majority voting of the eigenvalues of the K nearest neighbors. This method not only maintains the temporal continuity of the data but also reflects the actual changes in the distribution system's operation.

[0069] By effectively filling missing values, the integrity of the generated multidimensional feature set is ensured, providing a solid foundation for subsequent data analysis. Ensuring data integrity enables the model to effectively utilize all available data during training and inference, avoiding information loss due to missing data. This allows operations and maintenance personnel to make more accurate judgments based on more comprehensive historical data when analyzing and identifying distribution system faults, thereby improving the distribution system's fault warning capabilities and management capabilities.

[0070] During the specific implementation process, first, it is necessary to collect statistical information about missing values in the entire dataset. The system will automatically detect which data points are missing and record the timestamps of the missing values. Next, the system applies the K-nearest neighbor algorithm to interpolate the missing values. First, an appropriate K value is determined for each missing value (usually 3-5), and then the K nearest complete data points close to the time point of the missing value are identified. By calculating the characteristic values of these neighboring data points, the system uses a simple weighted average method to obtain an estimate of the missing value. To enhance the accuracy of interpolation, operations and maintenance personnel can also adjust the number of K and select different metrics (such as Euclidean distance or Manhattan distance) for interpolation. Finally, after interpolation is completed, all missing values will be filled with new estimated values, thereby ensuring data integrity and reliability in subsequent analysis.

[0071] Principal component analysis is used to reduce the dimensionality of the interpolated operating data, extracting the main electrical characteristics, environmental characteristics, and load characteristics to form a preliminary feature set. This preliminary feature set is then enhanced using a generative adversarial network to generate diverse feature samples.

[0072] Principal component analysis (PCA) is a commonly used dimensionality reduction technique that effectively reduces the dimensionality of data while preserving as much of the original information as possible. In data processing for power distribution systems, after the first two steps of preprocessing, the dataset may contain a large number of features, but not all features are equally important for fault analysis. PCA can identify the main factors affecting the operation of the distribution system and extract key components such as electrical characteristics, environmental characteristics, and load characteristics to form a preliminary feature set. Subsequently, using a generative adversarial network (GAN), data enhancement can be performed on this preliminary feature set, generating new data that is similar to the original data but has different feature samples, further enriching the diversity of the feature samples.

[0073] Through the dimensionality reduction process of principal component analysis, the feature set of the distribution system will become more concise and clear, retaining the most informative features, reducing redundant information and noise, and improving the training efficiency and analytical effectiveness of subsequent fault analysis models. In addition, using generative adversarial networks to generate diverse feature samples will further improve the model's generalization ability, enabling it to adapt to fault detection in different situations and greatly enhancing the system's ability to identify potential risks. This data augmentation method can effectively increase the number and representativeness of samples after the last principal component extraction, helping operations and maintenance personnel to more accurately predict faults.

[0074] In its implementation, the previously interpolated operational data is first normalized to ensure that the data range for each dimension of the feature falls within the same standard, thus avoiding bias due to different dimensions. Then, the PCA algorithm is used to calculate the feature covariance matrix and extract the eigenvalues and eigenvectors. The eigenvectors corresponding to the top few largest eigenvalues are selected to form a new feature set after dimensionality reduction. For example, if the analysis shows that the first five principal components can explain more than 90% of the variance, these five principal components will be used as the new feature set for subsequent analysis. To enhance the initial feature set, a generative adversarial network is constructed, consisting of a generator and a discriminator. The generator is responsible for generating samples that are similar to the original feature set but have different characteristics, while the discriminator evaluates the authenticity of the generated samples. After repeated training, the GAN will be able to generate a variety of feature samples, thereby increasing the diversity and richness of the feature set. Ultimately, after combining PCA dimensionality reduction with the generative adversarial network, the system outputs an enhanced feature set for subsequent fault analysis and model training.

[0075] Using a random forest algorithm to evaluate the contribution of each feature in the feature sample to fault prediction, and generating a feature importance score as a feature weight;

[0076] The random forest algorithm is an ensemble learning method that improves prediction accuracy and stability by constructing multiple decision trees. In power distribution system fault analysis, the random forest algorithm can be used to evaluate feature samples and calculate the contribution of each feature to fault prediction. The algorithm analyzes the splitting process of each tree to generate a feature importance score, reflecting the relative importance of each feature in fault prediction. This score provides a basis for subsequent feature selection and weighting, ensuring that the final model focuses on the most predictive features.

[0077] Generating feature importance scores effectively screens multidimensional features, helping the system focus on those with the greatest impact on fault analysis and improving the model's interpretability and predictive capabilities. This not only reduces model complexity and improves computational efficiency, but also helps operations personnel better understand and utilize these features in real-world operations. This feature selection process also provides a basis for subsequent model optimization, ensuring that operations decisions are based on well-defined features, enabling more timely and accurate detection and resolution of potential faults.

[0078] During the implementation, the previously generated feature samples are first organized into a feature matrix to ensure that the fault conditions corresponding to each feature are clearly labeled. Next, the feature matrix is trained using the random forest algorithm. Random forest constructs multiple decision trees by randomly extracting samples and features, and integrates the output results of each tree. By analyzing the features that each tree relies on when making a judgment, the frequency of use of each feature and its influence on the decision outcome are calculated. Finally, the importance scores of these features are aggregated to generate an overall feature importance ranking, which visualizes the contribution of the features to fault prediction. Based on these scores, operations and maintenance personnel can select the most important features for subsequent analysis to ensure that the model's predictive effect is maximized. During this process, cross-validation and other techniques can also be used to evaluate the robustness of the model and ensure the reliability of the feature importance scores.

[0079] The feature weights are used to weight each feature, and the weighted features are fused through a multi-layer perceptron to form a comprehensive multi-dimensional feature set. The multi-dimensional feature set can capture the real-time operating status of the power distribution system and reflect the relationship between electrical characteristics, environmental characteristics and load characteristics.

[0080] The generation of feature weights assigns different levels of importance to each feature. Fusion of these weighted features forms a comprehensive, multidimensional feature set that fully reflects the real-time operating status of the distribution system. This process utilizes a multilayer perceptron (MLP) to achieve feature integration and fusion. By nonlinearly mapping weighted features across multiple neuron layers, the MLP effectively captures the complex relationships between electrical, environmental, and load characteristics, enabling more comprehensive state representation and fault detection capabilities.

[0081] The multidimensional feature set generated through weighted fusion will provide richer information for subsequent fault analysis and prediction. This comprehensive multidimensional feature set not only improves the accuracy and stability of the analysis model but also helps operations and maintenance personnel gain clearer decision-making insights when handling complex power distribution issues. Furthermore, the construction of this comprehensive feature set lays a solid foundation for intelligent monitoring and management of power distribution systems, driving their development towards greater intelligence.

[0082] In the implementation, the previously calculated feature weights are first applied to each feature sample, resulting in a weighted feature set by weighting the feature values. Next, this weighted feature set is input into a multilayer perceptron, which typically consists of an input layer, multiple hidden layers, and an output layer. The input layer receives the weighted feature vectors, which are then transformed nonlinearly by multiple hidden layers to gradually extract higher-level feature relationships. For example, using the Reluctant Unit (ReLU) activation function in each hidden layer ensures that the model can learn complex nonlinear relationships. Ultimately, the output layer generates a comprehensive feature vector reflecting the various operating states of the distribution system and their interrelationships. For example, the output may include multidimensional information such as voltage stability, load balancing, and environmental impacts, providing a comprehensive foundation for subsequent fault analysis and real-time monitoring. This approach enables more accurate visualization of distribution system status and fault diagnosis, ensuring efficient operation and maintenance management.

[0083] S203, using a convolutional neural network to extract potential fault features from the multidimensional feature set and generate a fault feature vector, wherein the fault features include at least: load mutation, harmonic interference, and overvoltage;

[0084] In power distribution system fault analysis, extracting potential fault features using convolutional neural networks (CNNs) is a key step. This method automatically identifies characteristic patterns associated with faults through deep learning of multidimensional feature sets. Through a combination of multiple convolutional and pooling layers, CNNs effectively capture both local and global features in the signal, generating a fault feature vector. This fault feature vector not only includes features that directly reflect the fault state, such as sudden load changes, harmonic interference, and overvoltage, but also reveals deeper fault information through learned feature representations. This process enables the system to quickly and accurately identify potential fault features in complex operating environments, providing a critical basis for subsequent fault analysis and decision-making.

[0085] The process of extracting fault features through convolutional neural networks has significantly improved the intelligence of distribution system fault analysis. First, CNNs can automatically learn and extract features, reducing the need for manual feature engineering and improving analysis efficiency. Second, the generated fault feature vectors provide a rich information foundation for subsequent time series analysis and fault type identification, enabling the system to more accurately determine the fault type and severity. This approach not only improves the accuracy of fault detection but also enhances the system's real-time response capabilities, helping operations and maintenance personnel take timely measures to reduce the risks and losses caused by failures.

[0086] Specifically, a hierarchical convolutional neural network can be constructed, whose structure consists of multiple convolutional layers, pooling layers and fully connected layers.

[0087] When building a hierarchical convolutional neural network, the network's structural design is crucial. The network consists of multiple convolutional layers, pooling layers, and fully connected layers, and is divided into a local feature extraction module, a feature abstraction module, and a high-level feature synthesis module. The local feature extraction module uses multiple convolutional layers to extract local features from a multidimensional feature set and identify changes in the signal waveform. The feature abstraction module gradually reduces the dimensionality of the feature map to extract higher-level features, helping the model identify trends and periodicity in the signal. Finally, the high-level feature synthesis module converts the spatial dimensions of all high-level feature maps into one-dimensional vectors, forming a comprehensive feature vector, which lays the foundation for subsequent generation of fault feature vectors.

[0088] By constructing a hierarchical convolutional neural network, the system can effectively extract useful information from complex, multidimensional feature sets. This layered feature extraction approach enables the network to gradually learn feature representations from local to global perspectives, enhancing the model's expressiveness and generalization capabilities. The resulting fault feature vector provides a more precise basis for subsequent fault analysis, improving the accuracy and efficiency of fault detection.

[0089] In the specific implementation process, the appropriate convolution kernel size and stride are first selected to ensure that the network can effectively capture subtle changes in the signal. For example, using a small 3x3 convolution kernel for local feature extraction can identify transient mutations and waveform distortions in current and voltage signals. Next, a residual connection is applied after each convolution layer to add the local feature map output by the convolution layer to the input features at the element level to ensure effective transmission of information. Then, entering the feature abstraction module, continue to use the convolution layer and the maximum pooling layer to gradually reduce the dimension of the feature map to extract high-level features such as harmonic components and periodic fluctuation patterns. Finally, a global average pooling layer is used to convert the spatial dimensions of all high-level feature maps into one-dimensional vectors to form a comprehensive feature vector, which is passed to the fully connected layer to further learn the relationship between the features and finally generate a fault feature vector. The network is divided into three main modules:

[0090] Local feature extraction module: Through multiple convolutional layers, small convolution kernels are used to extract local features from the multidimensional feature set to obtain local feature maps to identify basic signal waveform changes. Signal waveform changes include instantaneous mutations and waveform distortions in current and voltage. The convolution operation in this layer is designed to capture subtle changes in the signal.

[0091] The local feature extraction module is the first step in a convolutional neural network. It primarily extracts local features from the signal through multiple convolutional layers. This module uses a small convolution kernel to convolve the input multidimensional feature set, effectively identifying transient mutations and waveform distortions in current and voltage signals. This process involves sliding the convolution kernel across a local feature map, extracting subtle changes in the signal and forming a local feature map, which provides the foundation for subsequent feature abstraction and synthesis.

[0092] The effectiveness of the local feature extraction module directly impacts the quality of subsequent feature abstraction and synthesis. By accurately capturing subtle changes in the signal, the system can better identify potential fault signatures. This process not only improves the accuracy of fault feature extraction but also provides a richer information foundation for subsequent fault analysis, enhancing the system's fault detection capabilities.

[0093] In the specific implementation, an appropriately small convolution kernel, such as a 3x3 or 5x5 kernel, is first selected and multiple convolution operations are performed. After each convolution, an activation function (such as ReLU) is used to perform a nonlinear transformation on the output to enhance the model's expressiveness. Next, a sliding window technique is used to perform convolution on the feature map, gradually extracting local features from the current and voltage signals. By stacking multiple layers of convolution, the system can gradually construct more complex local feature maps, ultimately forming a set of local feature maps that reflect the instantaneous changes and waveform distortion of the signal. These local feature maps will serve as input to the subsequent feature abstraction module to ensure the accuracy and effectiveness of subsequent analysis.

[0094] After each convolution layer is convolved, a residual connection is used to add the local feature map output by the convolution layer to the features in the multidimensional feature set input to the convolution layer at the element level to obtain a residual connection feature map;

[0095] Residual connections are an effective network architecture designed to address the vanishing gradient problem in deep network training. After each convolutional layer undergoes convolution, residual connections perform element-wise addition of the local feature map output by that convolutional layer with the input features to form a residual connection feature map. This process enables the network to learn the differences between input and output features, enhancing information transfer efficiency and ensuring effective feature retention.

[0096] By introducing residual connections, the network can better capture important features in the signal, avoiding the problem of gradual information loss in multi-layer networks. This structural design not only improves network training efficiency but also enhances the model's expressiveness, enabling the system to more accurately identify potential fault characteristics. The resulting residual connection feature map provides a richer information foundation for subsequent feature abstraction and synthesis.

[0097] In the specific implementation process, a convolution operation is first performed after each convolutional layer to obtain a local feature map. Next, this local feature map is element-wise added to the input features to form a residual connection feature map. To ensure dimensionality consistency, it may be necessary to perform an appropriate linear transformation (such as a 1x1 convolution) on the input features to match the dimensions of the local feature map. In this way, the residual connection feature map can effectively retain important information in the input features while introducing new local features, enhancing the network's learning ability. Ultimately, the residual connection feature map will serve as the input of the feature abstraction module to ensure the accuracy and effectiveness of subsequent analysis.

[0098] Feature Abstraction Module: Continuing to use convolutional layers and max pooling layers, the network gradually reduces the dimensionality of the residual connection feature map, enabling it to abstract high-level features from local features to obtain a high-level feature map. These high-level features include harmonic components and periodic fluctuation patterns. This module uses larger convolution kernels to help the model identify trends and periodic features in the signal.

[0099] The feature abstraction module's primary task is to extract higher-level features from local features. By continuing to use convolutional and max-pooling layers to gradually reduce the dimensionality of the residual connection feature map, the system is able to identify harmonic components and periodic fluctuations in the signal. This process, by using larger convolution kernels, helps the model capture trends and periodicity in the signal, thereby forming a high-level feature map that provides the foundation for subsequent feature synthesis.

[0100] The effectiveness of the feature abstraction module directly impacts the quality of high-level feature extraction. By abstracting high-level features from local features, the system can better identify complex patterns and trends in the signal. This process not only improves the accuracy of fault feature extraction but also provides a richer information foundation for subsequent fault analysis, enhancing the system's fault detection capabilities.

[0101] In its implementation, the system first selects an appropriate convolution kernel size (such as 5x5 or 7x7) and performs a convolution operation on the residual connection feature map to extract high-level features. Next, a maximum pooling layer is used to downsample the convolution output, reducing the dimensionality of the feature map and reducing computational complexity. By stacking multiple layers of convolution and pooling, the system can gradually construct more complex high-level feature maps, ultimately forming a set of high-level feature maps that reflect the signal's harmonic components and periodic fluctuation patterns. These high-level feature maps serve as input to the high-level feature synthesis module, ensuring the accuracy and effectiveness of subsequent analysis.

[0102] High-level feature synthesis module: Uses a global average pooling layer to convert the spatial dimensions of all high-level feature maps into one-dimensional vectors to retain the important information of each high-level feature map and form a comprehensive feature vector;

[0103] The main function of the global average pooling layer is to compress the spatial dimensions of each high-level feature map into a single value that represents the overall characteristics of that feature map. By averaging each feature map, the network can effectively extract the most representative feature information, reducing the dimensionality of the feature map while retaining important features. This method not only reduces computational complexity but also effectively prevents overfitting by reducing the number of model parameters.

[0104] Through global average pooling, the model can better focus on important information in the feature graph, rather than being distracted by local noise or irrelevant details. This processing method enables the model to more accurately capture key features in the signal during subsequent feature learning, thereby improving the accuracy and robustness of fault detection. For example, in power systems, fault characteristics such as sudden load changes, harmonic interference, and overvoltage can be effectively extracted and represented in this way, providing a reliable foundation for subsequent fault classification.

[0105] When implementing the global average pooling layer, you first need to input high-level feature maps into the layer. Assume that there are multiple high-level feature maps, and the size of each feature map is H*W, where H and W represent the height and width of the feature map, respectively. The global average pooling layer averages all pixel values in each feature map and outputs a one-dimensional vector with a length equal to the number of feature maps. For example, if there are 10 feature maps, after global average pooling, a vector of length 10 will be obtained. Each element in this vector corresponds to the average value of a feature map, representing the overall characteristics of the feature map. Next, this one-dimensional vector will be passed as input to the fully connected layer for further feature learning and classification.

[0106] The comprehensive feature vector is passed to the fully connected layer to further learn the relationship between features. The output vector generated by the softmax activation function is used as the fault feature vector. Each dimension of the fault feature vector corresponds to a different fault feature, where the fault features include at least: load mutation, harmonic interference and overvoltage.

[0107] The purpose of passing the comprehensive feature vector to the fully connected layer is to further learn the complex relationships between features. By weighting and biasing the input features, the fully connected layer can capture nonlinear relationships between features. After applying the softmax activation function, the output vector is converted into a probability distribution, representing the likelihood of each fault feature. This approach enables the model to effectively distinguish between multiple fault types, thereby improving fault detection accuracy.

[0108] Through processing at the fully connected layer, the model not only identifies fault signatures but also quantifies the probability of each fault signature occurring. This probability distribution provides an important basis for subsequent decision-making, enabling the system to make more accurate fault judgments in practical applications. For example, in a power distribution system, if the fault signature vector output by the model indicates a higher probability of a sudden load change than other fault signatures, the system can prioritize checking equipment related to the load change, thereby improving the efficiency of the fault response.

[0109] In practice, the comprehensive feature vector is input into one or more fully connected layers. Each fully connected layer consists of multiple neurons, each connected to each element of the input feature vector, performing a weighted sum and adding a bias. After processing with an activation function (such as ReLU), the output is passed to the next fully connected layer, until the final layer. In the final layer, a softmax activation function is applied to the output vector to generate a probability distribution for each fault feature. For example, assuming there are three fault features: load mutation, harmonic interference, and overvoltage, after softmax processing, the output might be [0.7, 0.2, 0.1], indicating a 70% probability of load mutation, a 20% probability of harmonic interference, and a 10% probability of overvoltage. This output provides important information for subsequent fault diagnosis and decision-making.

[0110] S204, performing time series analysis on the fault feature vector using a recurrent neural network to identify possible fault types in the current power distribution system, wherein the fault types include short circuit faults, ground faults, and equipment aging faults;

[0111] When using a recurrent neural network (RNN) to perform time series analysis on fault feature vectors, the fault feature vectors are first arranged in chronological order to form a time series dataset. This process ensures that the network can learn the temporal characteristics of the data, thereby capturing potential fault modes. At this stage, RNNs, particularly long short-term memory (LSTM) networks, are selected to process this time series data. Through its unique gating mechanism, LSTMs can effectively memorize information over long time spans, overcoming the vanishing or exploding gradient problem that traditional RNNs often encounter when learning long sequences. Therefore, by inputting the fault feature vectors at the current moment and several previous time steps, the LSTM network outputs a probability distribution representing the possible fault types and their corresponding probabilities in the current distribution system.

[0112] This process enables the distribution system to detect and identify fault types, such as short circuits, ground faults, and equipment aging, in real time through time-series analysis of fault characteristics. Timely fault identification significantly improves the safety and stability of the distribution system, reduces potential losses, lowers maintenance costs, and provides decision support for operations and maintenance personnel. Furthermore, based on probabilistic output, operations and maintenance personnel can prioritize faults based on their likelihood of occurrence, effectively allocating resources and rapidly responding to issues.

[0113] Specifically, the fault feature vectors can be arranged in chronological order to form a time series dataset;

[0114] In this step, fault feature vectors are sorted according to their acquisition timestamps to form a time series dataset. Each sample in a time series dataset contains the fault feature vector at the current time point and the feature vectors at the previous time points, thereby demonstrating the dynamic changes in fault features over time. This structure enables subsequent models to capture temporal dependencies, which is particularly important when fault patterns have specific temporal characteristics.

[0115] The significance of this time series structure lies in the system's ability to identify potential fault trends and changing patterns, thereby improving the accuracy of fault diagnosis. The characteristics of each time point are not only independent but also closely related to the previous state, forming a context-rich model foundation, which is crucial for dynamic monitoring of power distribution systems.

[0116] In practical applications, this data can be integrated with other fault signatures by recording the current and voltage changes of the device over specific time intervals. For example, a data collector can collect current and voltage signals once per second, extract the fault signatures for each acquisition, and store these signatures as a time series. When storing these signatures, the corresponding timestamps can be additionally saved to ensure that subsequent analysis can be interpolated according to the time sequence and input into the RNN model. In this way, the system provides a rich information foundation for fault analysis.

[0117] A pre-trained multi-layer long short-term memory network is used to perform time series analysis on the time series data set, wherein the input is the fault feature vector at the current moment and the fault feature vectors of the previous few time steps in the time series data set, and the output is a probability distribution representing the possible fault types and their probabilities in the current power distribution system. In addition, the structure of the multi-layer long short-term memory network includes: an input layer, multiple LSTM layers and a fully connected layer. The input layer receives time series data, the LSTM layer is responsible for extracting time series features, and the final fully connected layer is used to output the probability distribution of the fault type.

[0118] In this step, a multi-layer long short-term memory (LSTM) network is constructed and trained. This network is designed to analyze the generated time series dataset. The input layer receives the fault feature vector at the current moment and the feature vectors of the previous time steps, forming an input sequence containing multiple timestamp information. The LSTM layer has memory cells that effectively capture long-term dependencies in the time series and extract the temporal features in the sequence. Finally, after processing through multiple LSTM layers, the output layer generates a probability distribution through a fully connected layer, representing the possible fault types and probabilities in the current distribution system.

[0119] The significance of this analysis step lies in its ability to promptly identify and quantify possible fault types within the distribution system, such as short circuits, ground faults, and equipment aging, through deep time series feature extraction. By outputting probabilities, operations and maintenance personnel can prioritize high-risk faults based on their likelihood of occurrence, thereby improving the stability and reliability of the distribution system and avoiding significant economic losses or safety incidents.

[0120] During the implementation, the time series data is first preprocessed to ensure data standardization and cleanliness, removing any missing or outliers. Next, a multi-layered LSTM network structure is constructed, typically consisting of an input layer, several LSTM layers, and an output fully-connected layer. The input layer receives the constructed time series data, and the LSTM layer extracts temporal dynamic characteristics. Each LSTM layer processes the input sequence and generates a hidden state. This information is ultimately passed to the fully-connected layer, which generates a probability distribution for the fault type based on the current state. For example, suppose the LSTM network output indicates a 70% probability of a short circuit fault, a 20% probability of a ground fault, and a 10% probability of an equipment aging fault. This output helps maintenance teams quickly identify potential problems and make decisions, optimizing the troubleshooting process.

[0121] In step S205 , the identified fault type is spatially located based on the graph neural network in combination with the topology and electrical model of the power distribution system, and the specific node where the fault occurs and the corresponding fault severity level are output.

[0122] By combining the topology and electrical model of the distribution system and using graph neural networks to spatially locate identified fault types, we can effectively understand and analyze fault points in complex power networks. This process involves constructing a topological diagram of the system, identifying the nodes of each device and the electrical connections between them. The design of the graph neural network enables information propagation and aggregation between nodes on the topological diagram, thereby extracting fault information and assessing its severity. This approach not only improves the accuracy of fault location but also provides a scientific basis for subsequent troubleshooting, ensuring the safe and stable operation of the distribution system.

[0123] The significance of this method lies in improving the efficiency and accuracy of distribution system fault diagnosis. By spatially locating the fault type, operations and maintenance personnel can quickly identify the specific node where the fault occurred and take targeted measures to reduce power outage duration and economic losses. Furthermore, the analytical capabilities of graph neural networks enable managers to more comprehensively understand the system's operating status, supporting the optimization and intelligentization of distribution networks and driving the power industry towards more efficient and safe development.

[0124] Specifically, a topology diagram of the power distribution system may be constructed, wherein nodes of the topology diagram represent various devices in the power distribution system, and edges represent electrical connection relationships between the devices.

[0125] Constructing a topological map of a power distribution system is a crucial foundation for network fault diagnosis. A topological map consists of nodes and edges. Nodes represent individual devices in the distribution system (such as transformers, switches, and meters), while edges represent the electrical connections between these devices. By assigning unique identifiers and characteristic information to each device, the topological map intuitively reflects the overall structure and electrical characteristics of the system, providing the necessary framework for subsequent fault location analysis.

[0126] The significance of this process lies in establishing an effective visualization model for fault analysis, helping operations personnel and engineers quickly understand the overall layout of the distribution network and the status of key equipment. This structured representation of information not only facilitates subsequent analysis but also enables the system to respond quickly when faults occur, improving overall power system efficiency.

[0127] When constructing a topology map, each device in the distribution system must first be identified. Each device is assigned a unique node identifier, and its electrical characteristics (such as rated voltage and power) and environmental characteristics (such as location and temperature) are recorded. Next, the electrical connections between devices are studied, and weights are defined for each edge. These weights are typically related to the device's power capacity, load type, or electrical parameters, reflecting the degree of electrical influence between devices. Finally, an adjacency matrix is used to represent the connections between nodes, while a feature matrix contains the electrical and environmental characteristics recorded for each node. This creates a topological model of the system, laying the foundation for in-depth data analysis and fault location. For example, consider a distribution system with three transformers and several transmission lines. After constructing the topology map, each transformer is considered a node, and the connecting lines are considered edges, forming a complete network structure that facilitates subsequent information processing.

[0128] The steps of constructing the topology map include:

[0129] Assign a unique node identifier to each device and record its electrical and environmental characteristics;

[0130] When constructing a topology map, it's crucial to assign a unique node identifier to each device. This identifier ensures accurate identification of each device during fault location and fault type identification. In addition to the identifier, recording the device's electrical characteristics (such as rated voltage, current, and power factor) and environmental characteristics (such as installation location and operating temperature) is also crucial. This information aids subsequent fault analysis and decision support, enhancing the intelligence of the power system.

[0131] The primary purpose of this step is to provide sufficient information support for fault location. By recording the key parameters of each device, the fault identification system can more accurately assess and determine the device's operating status, thereby more effectively identifying the type and severity of the fault. This is crucial for improving the safety, reliability, and efficiency of power distribution system operations and maintenance.

[0132] When assigning identifiers to devices, you can use a simple coding scheme, such as combining the device model with a serial number, to ensure uniqueness. For example, transformer 1's identifier is "TRANS_001," and transformer 2's is "TRANS_002." Simultaneously, create a database or data sheet to record the electrical and environmental characteristics of each device. Electrical characteristics can include rated current, rated voltage, and power factor, while environmental characteristics can record the device's installation location and surrounding environmental conditions (such as humidity and temperature). This approach creates a comprehensive device information database, laying a solid foundation for fault detection and location.

[0133] Define edge weights based on the electrical connection relationships of the power distribution system. The weights are determined based on the electrical characteristics of the devices and reflect the electrical impact between the devices.

[0134] Defining edge weights in a topology graph is crucial for understanding the electrical characteristics of the power distribution system and the interactions between devices. Weights can be considered as indicators of the strength of the electrical connection between two devices and are typically assigned based on the device's electrical parameters. For example, factors such as load handling capacity, short-circuit current, and device capacity can more comprehensively describe the electrical impact between different devices, providing in-depth information for subsequent fault analysis.

[0135] This step enhances the effectiveness and accuracy of fault analysis by defining weights. The degree of electrical influence between different devices is directly related to the propagation and development of faults. Therefore, appropriate weights can help the fault location system better simulate and analyze the impact path of faults, improving its response speed and processing efficiency.

[0136] When defining edge weights, they can be quantified based on the device's power capacity, current load, and electrical characteristics. For example, for a connection between a transformer with a high power rating and a load, the edge weight can be set to a large value (such as 1.0), while for a relatively small device, it might be set to 0.5. This constructs a weighted adjacency matrix that effectively represents the electrical impact between devices. Weight design can be adjusted based on historical operating data or expert experience to ensure its rationality and reliability.

[0137] Representing a topological graph as an adjacency matrix and a characteristic matrix, wherein the adjacency matrix represents the connection relationship between nodes and the characteristic matrix contains the electrical and environmental characteristics of each node;

[0138] The mathematical representation of the topology graph is an important foundation for subsequent analysis. The adjacency matrix represents the connectivity between nodes in the network. The matrix elements reflect the degree of direct connectivity and weight between devices. The feature matrix records the electrical and environmental characteristics of each node device, providing the detailed information needed for decision support and analysis. This data structure provides a direct input foundation for graph neural networks, facilitating subsequent fault analysis and location.

[0139] The significance of this step lies in significantly improving the ability to process large-scale distribution system data through the structured representation of adjacency and feature matrices. The graph neural network model enables efficient computation of these matrices, enabling rapid fault detection, spatial localization, and assessment of equipment health, thereby optimizing maintenance strategies and decision-making processes.

[0140] First, when constructing the adjacency matrix, for each pair of devices, if there is a direct connection, the corresponding matrix position is filled with a weight value, and if there is no connection, it is filled with zero. For example, for a power distribution system with four devices, its adjacency matrix can be a 4x4 matrix, with rows and columns representing the four devices, and the intersection positions filled with electrical connection weight values. The feature matrix is established as a row recording the electrical characteristics and environmental characteristics of each device, which may include voltage, current, temperature, humidity, etc. Ultimately, these two matrices can be input into the graph neural network for subsequent analysis and fault diagnosis.

[0141] The identified fault type is mapped to the nodes of the topology graph, and the mapped topology graph is input into the pre-trained graph neural network to output the specific node where the fault occurred and the corresponding fault severity level.

[0142] This step combines fault type information with the topology map, mapping the identified fault type to the corresponding device node. By inputting a pre-trained graph neural network, it leverages its powerful graph structure learning capabilities to conduct in-depth fault analysis, outputting the specific location of each fault node and its severity level. This process is a critical step in fault detection and location, ensuring efficient and accurate information flow.

[0143] The purpose of this process is to effectively reduce the dimension of complex fault information through the computing power of graph neural networks and accurately map it to specific equipment nodes, greatly improving the efficiency and accuracy of fault location. It provides a clear goal for operation and maintenance personnel, so that they can take timely measures to repair the fault and ensure the safe and stable operation of the distribution system.

[0144] During the implementation process, it is first necessary to match the identified fault types (such as short circuit faults, ground faults, etc.) with the corresponding nodes in the distribution system topology map to ensure that each fault information can accurately point to a specific device. Next, the topology map containing the fault information is fed into the pre-trained graph neural network model as input data. By learning the connection relationship between nodes, the model can automatically extract features and determine the location and severity of each fault. For example, if a transformer is marked as having a short circuit fault, the graph neural network will process the adjacency matrix and feature matrix in the topology map, and after multiple layers of graph convolution operations, output the fault level of the transformer (such as "high voltage short circuit"). This result will be directly fed back to the operation and maintenance personnel for rapid response and processing.

[0145] It can be seen that the operation-related data of the distribution system are collected in real time through a variety of sensors and intelligent metering equipment; the collected operation-related data are preprocessed to remove noise and interference signals, and a multi-dimensional feature set of the real-time operation status of the distribution system is established; the convolutional neural network is used to extract potential fault features from the multi-dimensional feature set and generate a fault feature vector; the fault feature vector is analyzed in time series using a recurrent neural network to identify the possible fault types in the current distribution system; combined with the topological structure and electrical model of the distribution system, the identified fault type is spatially located based on the graph neural network, and the specific node where the fault occurs and the corresponding fault severity level are output, which can significantly improve the fault detection capability and response speed of the distribution system and promote the development of the distribution system towards intelligence and automation.

[0146] Another embodiment of the present invention provides a power distribution system fault analysis system, see Figure 3 , the system may include:

[0147] The acquisition module 301 is used to collect the operation-related data of the power distribution system in real time through various sensors and intelligent metering devices;

[0148] Processing module 302, configured to pre-process the collected operation-related data to remove noise and interference signals and establish a multi-dimensional feature set of the real-time operation status of the power distribution system, wherein the multi-dimensional feature set includes electrical characteristics, environmental characteristics, and load characteristics;

[0149] An extraction module 303 is configured to extract potential fault features from the multidimensional feature set using a convolutional neural network and generate a fault feature vector, wherein the fault features include at least load mutation, harmonic interference, and overvoltage;

[0150] Identification module 304, configured to perform time series analysis on the fault feature vector using a recurrent neural network to identify possible fault types in the current power distribution system, wherein the fault types include short circuit faults, ground faults, and equipment aging faults;

[0151] The output module 305 is used to combine the topology and electrical model of the power distribution system, spatially locate the identified fault type based on the graph neural network, and output the specific node where the fault occurred and the corresponding fault severity level.

[0152] It can be seen that the operation-related data of the distribution system are collected in real time through a variety of sensors and intelligent metering equipment; the collected operation-related data are preprocessed to remove noise and interference signals, and a multi-dimensional feature set of the real-time operation status of the distribution system is established; the convolutional neural network is used to extract potential fault features from the multi-dimensional feature set and generate a fault feature vector; the fault feature vector is analyzed in time series using a recurrent neural network to identify the possible fault types in the current distribution system; combined with the topological structure and electrical model of the distribution system, the identified fault type is spatially located based on the graph neural network, and the specific node where the fault occurs and the corresponding fault severity level are output, which can significantly improve the fault detection capability and response speed of the distribution system and promote the development of the distribution system towards intelligence and automation.

[0153] An embodiment of the present invention further provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps of any one of the above method embodiments when running.

[0154] Specifically, in this embodiment, the above-mentioned storage medium may be configured to store a computer program for performing the following steps:

[0155] S201, collects real-time operation data of the power distribution system through various sensors and smart metering equipment;

[0156] S202, preprocessing the collected operation-related data to remove noise and interference signals, and establishing a multi-dimensional feature set of the real-time operation status of the power distribution system, wherein the multi-dimensional feature set includes electrical characteristics, environmental characteristics, and load characteristics;

[0157] S203, using a convolutional neural network to extract potential fault features from the multidimensional feature set and generate a fault feature vector, wherein the fault features include at least: load mutation, harmonic interference, and overvoltage;

[0158] S204, performing time series analysis on the fault feature vector using a recurrent neural network to identify possible fault types in the current power distribution system, wherein the fault types include short circuit faults, ground faults, and equipment aging faults;

[0159] In step S205 , the identified fault type is spatially located based on the graph neural network in combination with the topology and electrical model of the power distribution system, and the specific node where the fault occurs and the corresponding fault severity level are output.

[0160] It can be seen that the operation-related data of the distribution system are collected in real time through a variety of sensors and intelligent metering equipment; the collected operation-related data are preprocessed to remove noise and interference signals, and a multi-dimensional feature set of the real-time operation status of the distribution system is established; the convolutional neural network is used to extract potential fault features from the multi-dimensional feature set and generate a fault feature vector; the fault feature vector is analyzed in time series using a recurrent neural network to identify the possible fault types in the current distribution system; combined with the topological structure and electrical model of the distribution system, the identified fault type is spatially located based on the graph neural network, and the specific node where the fault occurs and the corresponding fault severity level are output, which can significantly improve the fault detection capability and response speed of the distribution system and promote the development of the distribution system towards intelligence and automation.

[0161] An embodiment of the present invention further provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any one of the above method embodiments.

[0162] Specifically, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor, and the input / output device is connected to the processor.

[0163] Specifically, in this embodiment, the processor may be configured to execute the following steps through a computer program:

[0164] S201, collects real-time operation data of the power distribution system through various sensors and smart metering equipment;

[0165] S202, preprocessing the collected operation-related data to remove noise and interference signals, and establishing a multi-dimensional feature set of the real-time operation status of the power distribution system, wherein the multi-dimensional feature set includes electrical characteristics, environmental characteristics, and load characteristics;

[0166] S203, using a convolutional neural network to extract potential fault features from the multidimensional feature set and generate a fault feature vector, wherein the fault features include at least: load mutation, harmonic interference, and overvoltage;

[0167] S204, performing time series analysis on the fault feature vector using a recurrent neural network to identify possible fault types in the current power distribution system, wherein the fault types include short circuit faults, ground faults, and equipment aging faults;

[0168] In step S205 , the identified fault type is spatially located based on the graph neural network in combination with the topology and electrical model of the power distribution system, and the specific node where the fault occurs and the corresponding fault severity level are output.

[0169] It can be seen that the operation-related data of the distribution system are collected in real time through a variety of sensors and intelligent metering equipment; the collected operation-related data are preprocessed to remove noise and interference signals, and a multi-dimensional feature set of the real-time operation status of the distribution system is established; the convolutional neural network is used to extract potential fault features from the multi-dimensional feature set and generate a fault feature vector; the fault feature vector is analyzed in time series using a recurrent neural network to identify the possible fault types in the current distribution system; combined with the topological structure and electrical model of the distribution system, the identified fault type is spatially located based on the graph neural network, and the specific node where the fault occurs and the corresponding fault severity level are output, which can significantly improve the fault detection capability and response speed of the distribution system and promote the development of the distribution system towards intelligence and automation.

[0170] The above describes in detail the structure, features and effects of the present invention based on the embodiments shown in the drawings. The above is only a preferred embodiment of the present invention, but the scope of implementation of the present invention is not limited to what is shown in the drawings. Any changes made in accordance with the concept of the present invention, or modifications to equivalent embodiments with equivalent changes, which do not exceed the spirit covered by the description and drawings, should be within the scope of protection of the present invention.

Claims

1. A method for analyzing power distribution system faults, characterized in that: The method comprises: Collect distribution system operation-related data in real time through various sensors and smart metering devices; Preprocessing the collected operation-related data to remove noise and interference signals and establish a multi-dimensional feature set of the real-time operation status of the power distribution system, wherein the multi-dimensional feature set includes electrical characteristics, environmental characteristics, and load characteristics; A convolutional neural network is used to extract potential fault features from the multidimensional feature set and generate a fault feature vector, wherein the fault features include at least: load mutation, harmonic interference and overvoltage; Using a recurrent neural network to perform time series analysis on the fault feature vectors to identify possible fault types in the current power distribution system, where the fault types include short circuit faults, ground faults, and equipment aging faults; Combining the topological structure and electrical model of the power distribution system, the identified fault type is spatially located based on the graph neural network, and the specific node where the fault occurs and the corresponding fault severity level are output.

2. The method according to claim 1, characterized in that The pre-processing of the collected operation-related data to remove noise and interference signals and establish a multi-dimensional feature set of the real-time operation status of the power distribution system includes: The current and voltage signals in the collected operating data are decomposed using wavelet transform to extract low-frequency and high-frequency components, and high-frequency noise is removed through threshold processing to retain the main features of the signal; For missing values that may appear during the data collection process, an interpolation method based on the K-nearest neighbor algorithm is used to fill them according to the characteristic values of similar time points to ensure the integrity of the operating data; Principal component analysis is used to reduce the dimensionality of the interpolated operating data, extracting the main electrical characteristics, environmental characteristics, and load characteristics to form a preliminary feature set. This preliminary feature set is then enhanced using a generative adversarial network to generate diverse feature samples. Using a random forest algorithm to evaluate the contribution of each feature in the feature sample to fault prediction, and generating a feature importance score as a feature weight; The feature weights are used to weight each feature, and the weighted features are fused through a multi-layer perceptron to form a comprehensive multi-dimensional feature set. The multi-dimensional feature set can capture the real-time operating status of the power distribution system and reflect the relationship between electrical characteristics, environmental characteristics and load characteristics.

3. The method according to claim 2, characterized in that The convolutional neural network is used to extract potential fault features from the multidimensional feature set and generate a fault feature vector, wherein the fault features include at least: load mutation, harmonic interference and overvoltage, including: Construct a hierarchical convolutional neural network whose structure consists of multiple convolutional layers, pooling layers and fully connected layers; the network is divided into three main modules: Local feature extraction module: Through multiple convolutional layers, small convolution kernels are used to extract local features from the multidimensional feature set to obtain local feature maps to identify basic signal waveform changes. Signal waveform changes include instantaneous mutations and waveform distortions in current and voltage. The convolution operation in this layer is designed to capture subtle changes in the signal. After each convolution layer is convolved, a residual connection is used to add the local feature map output by the convolution layer to the features in the multidimensional feature set input to the convolution layer at the element level to obtain a residual connection feature map; Feature Abstraction Module: Continuing to use convolutional layers and max pooling layers, the network gradually reduces the dimensionality of the residual connection feature map, enabling it to abstract high-level features from local features to obtain a high-level feature map. These high-level features include harmonic components and periodic fluctuation patterns. This module uses larger convolution kernels to help the model identify trends and periodic features in the signal. High-level feature synthesis module: Uses a global average pooling layer to convert the spatial dimensions of all high-level feature maps into one-dimensional vectors to retain the important information of each high-level feature map and form a comprehensive feature vector; The comprehensive feature vector is passed to the fully connected layer to further learn the relationship between features. The output vector generated by the softmax activation function is used as the fault feature vector. Each dimension of the fault feature vector corresponds to a different fault feature, where the fault features include at least: load mutation, harmonic interference and overvoltage.

4. The method according to claim 3, characterized in that The method of using a recurrent neural network to perform time series analysis on the fault feature vector to identify the possible fault types in the current power distribution system includes: Arrange the fault feature vectors in chronological order to form a time series data set; A pre-trained multi-layer long short-term memory network is used to perform time series analysis on the time series data set, wherein the input is the fault feature vector at the current moment and the fault feature vectors of the previous few time steps in the time series data set, and the output is a probability distribution representing the possible fault types and their probabilities in the current power distribution system. In addition, the structure of the multi-layer long short-term memory network includes: an input layer, multiple LSTM layers and a fully connected layer. The input layer receives time series data, the LSTM layer is responsible for extracting time series features, and the final fully connected layer is used to output the probability distribution of the fault type.

5. The method according to claim 4, characterized in that The method combines the topology and electrical model of the power distribution system and spatially locates the identified fault type based on the graph neural network, outputting the specific node where the fault occurred and the corresponding fault severity level, including: Construct a topology diagram of the power distribution system, where the nodes of the topology diagram represent the devices in the power distribution system, and the edges represent the electrical connections between the devices. The steps of constructing the topology diagram include: Assign a unique node identifier to each device and record its electrical and environmental characteristics; Define edge weights based on the electrical connection relationships of the power distribution system. The weights are determined based on the electrical characteristics of the devices and reflect the electrical impact between the devices. Representing a topological graph as an adjacency matrix and a characteristic matrix, wherein the adjacency matrix represents the connection relationship between nodes and the characteristic matrix contains the electrical and environmental characteristics of each node; The identified fault type is mapped to the nodes of the topology graph, and the mapped topology graph is input into the pre-trained graph neural network to output the specific node where the fault occurred and the corresponding fault severity level.

6. A power distribution system fault analysis system, characterized in that: The system comprises: The acquisition module is used to collect real-time operation-related data of the power distribution system through various sensors and intelligent metering devices; a processing module, configured to pre-process the collected operation-related data to remove noise and interference signals and establish a multi-dimensional feature set of the real-time operation status of the power distribution system, wherein the multi-dimensional feature set includes electrical characteristics, environmental characteristics, and load characteristics; An extraction module, configured to extract potential fault features from the multidimensional feature set using a convolutional neural network and generate a fault feature vector, wherein the fault features include at least load mutation, harmonic interference, and overvoltage; An identification module is used to perform time series analysis on the fault feature vector using a recurrent neural network to identify possible fault types in the current power distribution system, wherein the fault types include short circuit faults, ground faults, and equipment aging faults; The output module is used to combine the topology and electrical model of the distribution system, spatially locate the identified fault type based on the graph neural network, and output the specific node where the fault occurred and the corresponding fault severity level.

7. The system according to claim 6, characterized in that The processing module is specifically used to: The current and voltage signals in the collected operating data are decomposed using wavelet transform to extract low-frequency and high-frequency components, and high-frequency noise is removed through threshold processing to retain the main features of the signal. For missing values that may appear during the data collection process, an interpolation method based on the K-nearest neighbor algorithm is used to fill in the missing values according to the characteristic values of similar time points to ensure the integrity of the running data; Principal component analysis is used to reduce the dimensionality of the interpolated operating data, extracting the main electrical characteristics, environmental characteristics, and load characteristics to form a preliminary feature set. This preliminary feature set is then enhanced using a generative adversarial network to generate diverse feature samples. Using a random forest algorithm to evaluate the contribution of each feature in the feature sample to fault prediction, and generating a feature importance score as a feature weight; The feature weights are used to weight each feature, and the weighted features are fused through a multi-layer perceptron to form a comprehensive multi-dimensional feature set. The multi-dimensional feature set can capture the real-time operating status of the power distribution system and reflect the relationship between electrical characteristics, environmental characteristics and load characteristics.

8. The system according to claim 7, characterized in that The extraction module is specifically used to: Construct a hierarchical convolutional neural network whose structure consists of multiple convolutional layers, pooling layers and fully connected layers; the network is divided into three main modules: Local feature extraction module: Through multiple convolutional layers, small convolution kernels are used to extract local features from the multidimensional feature set to obtain local feature maps to identify basic signal waveform changes. Signal waveform changes include instantaneous mutations and waveform distortions in current and voltage. The convolution operation in this layer is designed to capture subtle changes in the signal. After each convolution layer is convolved, a residual connection is used to add the local feature map output by the convolution layer to the features in the multidimensional feature set input to the convolution layer at the element level to obtain a residual connection feature map; Feature Abstraction Module: Continuing to use convolutional layers and max pooling layers, the network gradually reduces the dimensionality of the residual connection feature map, enabling it to abstract high-level features from local features to obtain a high-level feature map. These high-level features include harmonic components and periodic fluctuation patterns. This module uses larger convolution kernels to help the model identify trends and periodic features in the signal. High-level feature synthesis module: Uses a global average pooling layer to convert the spatial dimensions of all high-level feature maps into one-dimensional vectors to retain the important information of each high-level feature map and form a comprehensive feature vector; The comprehensive feature vector is passed to the fully connected layer to further learn the relationship between features. The output vector generated by the softmax activation function is used as the fault feature vector. Each dimension of the fault feature vector corresponds to a different fault feature, where the fault features include at least: load mutation, harmonic interference and overvoltage.

9. A storage medium, characterized in that: The storage medium stores a computer program, wherein the computer program is configured to execute the method according to any one of claims 1 to 5 when run.

10. An electronic device comprising a memory and a processor, characterized in that: A computer program is stored in the memory, and the processor is configured to run the computer program to perform the method according to any one of claims 1 to 5.

Images