Power distribution equipment fault probability comprehensive quantitative evaluation method and system

Abstract

The invention discloses a power distribution equipment fault probability comprehensive quantitative evaluation method and system, relates to the technical field of power distribution equipment, and realizes intelligent management of power distribution equipment fault prediction and maintenance strategies by integrating a physical information neural network, abnormal score analysis, a space-time diagram neural network, maintenance decision calculation and the like. The accuracy and reliability of fault prediction are improved, the false alarm rate is reduced through probability output, and uncertainty quantization is supported; the sensitivity of anomaly detection is enhanced, early warning and priority ranking are facilitated, and the response speed of the system to an abnormal state is improved; the space and time dependency relationship is comprehensively modeled, the risk assessment and resource allocation are optimized, and the cascading fault risk is reduced; a data-driven maintenance strategy is generated, the maintenance efficiency and quality are improved, the cost is reduced, and the downtime is shortened; a dynamic risk portrait is generated through attribution analysis, real-time monitoring and decision making are supported, and system transparency is enhanced.

Description

Technical Field

[0001] This application relates to the field of power distribution equipment technology, specifically to a comprehensive quantitative assessment method and system for power distribution equipment failure probability. Background Technology

[0002] Traditional power distribution equipment fault assessment suffers from problems such as strong subjectivity, insufficient data utilization, single assessment dimensions, and lack of dynamic prediction capabilities. Therefore, it is crucial to construct a quantitative, dynamic, comprehensive, and interpretable intelligent assessment system to improve power supply reliability, optimize asset allocation, and reduce operation and maintenance costs.

[0003] Existing technologies, such as the invention application patent with publication number CN112507290A, disclose a method, device, and storage medium for predicting the failure probability of power distribution equipment. This aims to solve the problems of low accuracy, difficulty in quantification, and high cost in existing technologies. It includes: real-time acquisition of the failure category and failure factors of the power distribution equipment; processing the failure category and failure factors of the power distribution equipment using a pre-constructed multi-layer failure probability fusion model to obtain an initial predicted value of the failure probability of the power distribution equipment; and using a probability threshold to determine the validity of the initial predicted value of the failure probability to obtain the predicted failure probability result of the power distribution equipment.

[0004] Regarding the above-mentioned solutions, the inventors of this application have found that the above-mentioned technologies have at least the following technical problems: 1. Currently, there is a lack of processing using physical information neural networks, no generation of preliminary fault probability assessment data, a lack of integration of physical models and data-driven methods, and the inability to reduce false alarm rates, identify potential faults in advance, or reduce unexpected downtime by using probability output instead of point estimation; there is a lack of anomaly scoring analysis based on preliminary fault probability assessment data, the inability to derive quantitative anomaly scores from probability assessments, the inability to achieve fine-grained anomaly detection, the inability to enhance the system's sensitivity to abnormal states, and the inability to improve the comprehensiveness of detection and response speed.

[0005] 2. Currently, there is a lack of modeling spatial and temporal dependencies using spatiotemporal graph neural networks to generate risk propagation graph data. This fails to capture dynamic risk propagation in the device network, thus hindering the real-time performance and accuracy of risk assessment and preventing optimized resource allocation. Furthermore, there is a lack of maintenance decision-making calculations based on risk propagation graph data and maintenance resource data, and the inability to integrate multi-source data to generate optimal maintenance strategies. This fails to improve maintenance efficiency and quality, and cannot reduce maintenance costs and time. Finally, the lack of dynamic risk profiling to support real-time monitoring and decision-making fails to enhance system transparency and interpretability, and cannot assist maintenance personnel in rapid response. Summary of the Invention

[0006] To address the aforementioned technical shortcomings, the purpose of this application is to provide a comprehensive quantitative assessment method and system for the failure probability of power distribution equipment.

[0007] To solve the above-mentioned technical problems, this application adopts the following technical solution: In the first aspect, this application provides a comprehensive quantitative assessment method for the failure probability of power distribution equipment. The method includes the following steps: Step 1, preliminary failure probability analysis: Based on the pre-acquired power distribution equipment data, the data is processed through a physical information neural network to generate preliminary failure probability assessment data.

[0008] Step 2, Anomaly Score Data Analysis: Based on the preliminary fault probability assessment data, perform anomaly scoring analysis to obtain the original anomaly score and generate anomaly score data.

[0009] Step 3: Risk propagation graph data generation: Based on the anomaly score data and the pre-acquired device network topology data, spatial and temporal dependencies are modeled using a spatiotemporal graph neural network to generate risk propagation graph data.

[0010] Step 4: Maintenance strategy data generation: Based on the risk propagation map data and the pre-acquired maintenance resource data, maintenance decision calculations are performed to obtain the original maintenance action probability distribution and priority score, so as to generate the optimal maintenance strategy data.

[0011] Step 5: Dynamic Risk Profile Data Generation: Attribution analysis is performed based on the optimal maintenance strategy data to generate dynamic risk profile data.

[0012] Preferably, the pre-acquired power distribution equipment data includes static data, dynamic operation data, status monitoring data, and environmental data.

[0013] Preferably, the step of processing the pre-acquired power distribution equipment data through a physical information neural network to generate preliminary fault probability assessment data includes: performing data cleaning and normalization processing on the pre-acquired power distribution equipment data to obtain standardized power distribution equipment data; inputting the standardized power distribution equipment data into a pre-trained physical information neural network, and generating preliminary fault probability assessment data through forward propagation calculation of the physical information neural network; and incorporating physical mechanism constraints into the physical information neural network during training.

[0014] Preferably, the step of performing anomaly scoring analysis based on the preliminary fault probability assessment data to obtain an original anomaly score and generating anomaly score data includes: performing data fusion processing based on the preliminary fault probability assessment data and pre-acquired power distribution equipment data to obtain multimodal fusion data; inputting the multimodal fusion data into a pre-trained deep learning model for feature extraction and anomaly detection calculation to obtain an original anomaly score; and performing post-processing optimization on the original anomaly score to generate anomaly score data.

[0015] Preferably, the step of modeling spatial and temporal dependencies using a spatiotemporal graph neural network based on the anomaly score data and pre-acquired device network topology data to generate risk propagation graph data includes: performing graph structure construction processing based on the anomaly score data and pre-acquired device network topology data to generate spatiotemporal graph data; inputting the spatiotemporal graph data into a pre-trained spatiotemporal graph neural network model, processing spatial dependencies through a spatial graph convolutional layer to obtain a spatial feature matrix; further processing temporal dependencies through a temporal convolutional layer to obtain a spatiotemporal fusion feature matrix; and then performing output processing to calculate the mutual influence risk between devices using a risk scoring function to generate risk propagation graph data.

[0016] Preferably, the output processing, which calculates the mutual influence risk between devices using a risk scoring function to generate risk propagation map data, includes: A1, using a risk scoring function... The result is the The device and the first Risk propagation score of each device , and Represent the spatiotemporal fusion feature matrix, respectively. The device and the first Feature vectors of each device The function is represented as compressing the output to a probability range.

[0017] A2. Based on the risk propagation score, post-processing optimization is performed, including normalization and threshold filtering, to generate the final risk propagation map data, which includes the risk score matrix corresponding to all devices.

[0018] Preferably, the step of calculating the original maintenance action probability distribution and priority score based on the risk propagation map data and pre-acquired maintenance resource data to generate optimal maintenance strategy data includes: performing state vector construction processing based on the risk propagation map data and pre-acquired maintenance resource data to obtain a state representation vector; then inputting the state representation vector into a pre-trained deep reinforcement learning model, performing maintenance decision calculation through a policy network to obtain the original maintenance action probability distribution and priority score; and performing decision optimization processing on the original maintenance action probability distribution and priority score to generate optimal maintenance strategy data.

[0019] Preferably, the step of performing decision optimization processing on the original maintenance action probability distribution and priority score to generate optimal maintenance strategy data includes: the decision optimization processing selects the maintenance action with the highest probability based on the original maintenance action probability distribution as the final action; and performs discretization processing based on the priority score to determine the maintenance priority.

[0020] Preferably, the attribution analysis based on the optimal maintenance strategy data to generate dynamic risk profile data, displaying the main risk factors and their contributions, includes: performing data integration processing based on the optimal maintenance strategy data and the data from steps one, two, and three to obtain attribution analysis input data; inputting the attribution analysis input data into a pre-trained interpretable artificial intelligence model, performing feature importance analysis through the attribution calculation layer to obtain the original contribution score, performing normalization processing to obtain the standardized contribution score; and generating dynamic risk profile data based on the standardized contribution score, wherein the dynamic risk profile data includes the main risk factors and their corresponding contributions.

[0021] In its second aspect, this application provides a system for a comprehensive quantitative assessment method of power distribution equipment failure probability, comprising: a preliminary failure probability analysis module, which processes pre-acquired power distribution equipment data through a physical information neural network to generate preliminary failure probability assessment data.

[0022] The anomaly score data analysis module performs anomaly scoring analysis based on the preliminary fault probability assessment data to obtain the original anomaly score and generate anomaly score data.

[0023] The risk propagation graph data generation module, based on the anomaly score data and pre-acquired device network topology data, uses a spatiotemporal graph neural network to model spatial and temporal dependencies in order to generate risk propagation graph data.

[0024] The maintenance strategy data generation module, based on the risk propagation map data and the pre-acquired maintenance resource data, performs maintenance decision calculations to obtain the original maintenance action probability distribution and priority score, so as to generate the optimal maintenance strategy data;

[0025] The dynamic risk profile data generation module performs attribution analysis based on optimal maintenance strategy data to generate dynamic risk profile data.

[0026] The beneficial effects of this application are as follows: 1. The comprehensive quantitative assessment method and system for the probability of power distribution equipment failure provided in this application, by integrating physical information neural networks, anomaly scoring analysis, spatiotemporal graph neural networks, and maintenance decision calculation, realizes intelligent management of power distribution equipment failure prediction and maintenance strategies, improves the accuracy and reliability of failure prediction, reduces false alarm rate through probability output, and supports uncertainty quantification; enhances the sensitivity of anomaly detection, facilitates early warning and priority ranking, and improves the system's response speed to abnormal states; comprehensively models spatial and temporal dependencies, optimizes risk assessment and resource allocation, and reduces the risk of cascading failures; generates data-driven maintenance strategies, improves maintenance efficiency and quality, and reduces costs and downtime; generates dynamic risk profiles through attribution analysis, supports real-time monitoring and decision-making, and enhances system transparency.

[0027] 2. This application uses a physical information neural network for processing to generate preliminary fault probability assessment data. It integrates physical models and data-driven methods, and reduces the false alarm rate by using probability output instead of point estimation. It also provides uncertainty quantification, making the prediction results more reliable, thereby identifying potential faults in advance and reducing unexpected downtime. Based on the preliminary fault probability assessment data, anomaly scoring analysis is performed, and quantitative anomaly scores are derived from the probability assessment. This enables fine-grained anomaly detection, which facilitates priority ranking and early warning, enhances the system's sensitivity to abnormal states, and improves the comprehensiveness of detection and response speed.

[0028] 3. This application uses a spatiotemporal graph neural network to model spatial and temporal dependencies, generate risk propagation graph data, capture dynamic risk propagation in the equipment network, improve the real-time performance and accuracy of risk assessment, help identify risk hotspots and propagation paths, thereby optimizing resource allocation and reducing failures; based on risk propagation graph data and maintenance resource data, maintenance decision calculations are performed, and multi-source data are integrated to generate the optimal maintenance strategy, which improves maintenance efficiency and quality, reduces maintenance costs and time, and emphasizes probabilistic optimization and resource constraints.

[0029] 4. The dynamic risk profile in this application supports real-time monitoring and decision-making, and reveals the root causes through attribution analysis, thereby enhancing the transparency and interpretability of the system and assisting operation and maintenance personnel in responding quickly. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 This is a flowchart illustrating the implementation steps of the method described in this application.

[0032] Figure 2 This is a schematic diagram of the system structure connection of this application. Detailed Implementation

[0033] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0034] Please see Figure 1As shown, this application provides a comprehensive quantitative assessment method for the failure probability of power distribution equipment in the first aspect, including: Step 1, preliminary failure probability analysis: Based on the pre-acquired power distribution equipment data, the data is processed through a physical information neural network to generate preliminary failure probability assessment data.

[0035] In a specific instance, the pre-acquired power distribution equipment data includes static data, dynamic operation data, status monitoring data, and environmental data.

[0036] It should be noted that static data represents inherent parameters of the equipment, including equipment type, specifications, material properties (such as resistivity and tensile strength), and network topology; dynamic operating data includes generator data (such as power output, frequency, temperature, pressure, and efficiency), transformer data (such as load rate, input and output voltage, current, and operating status of switching equipment), and power distribution network data (such as line load, voltage level, line temperature, and fault indication); condition monitoring data are real-time monitoring indicators, including equipment temperature, pressure, flow rate, and fault indication; environmental data represents external environmental factors, including weather information, temperature, humidity, wind speed, and precipitation data.

[0037] In a specific example, the process of generating preliminary fault probability assessment data based on pre-acquired power distribution equipment data through a physical information neural network includes: performing data cleaning and normalization on the pre-acquired power distribution equipment data to obtain standardized power distribution equipment data; inputting the standardized power distribution equipment data into a pre-trained physical information neural network, and generating preliminary fault probability assessment data through forward propagation calculation of the physical information neural network; and incorporating physical mechanism constraints into the physical information neural network during training.

[0038] It should be noted that the physical information neural network incorporates physical mechanism constraints during training, which is achieved through a physical loss function, the mathematical expression of which is: ,in Represented as the total loss function, Represented as a data loss function, This is represented as a correction factor. It is the physical loss function.

[0039] Furthermore, the specific formula for the data loss function is as follows: ,in This represents the ID corresponding to the training sample. , This represents the total number of training samples. Represented as a neural network for the first... The predicted failure probability value for each training sample. Represented as the first The true fault labels of the training samples.

[0040] Furthermore, the specific formula for the physical loss function is as follows: ,in This represents the ID corresponding to the sample in the current training batch. , This represents the total number of samples in the current training batch; This represents the output of the neural network in the current training batch, specifically the [number]th [unit]. Power-related predicted values ​​for each sample. and These represent the nth and nth elements in the current training batch. Input voltage and current data for each sample. This represents the output of the neural network in the current training batch, specifically the [number]th [unit]. Temperature-related predicted values ​​for each sample It is represented as the resistivity of the device; the physical loss function ensures that the neural network incorporates physical mechanism constraints by penalizing the difference between the predicted value and the calculated value of the physical equation.

[0041] It should be noted that the total loss function represents the overall error that needs to be minimized during training; the data loss function is calculated based on the difference between the predicted value and the true fault label; the physical loss function calculates the residual based on the physical equations of the power distribution equipment to ensure that the output conforms to physical laws; and the correction coefficient is used to balance the contributions of data loss and physical loss, and Power-related predicted values ​​represent the correlation between failure probability and power; temperature-related predicted values ​​represent the correlation between failure probability and temperature; the resistivity of the equipment is determined based on the material properties of the equipment, for example, for copper wires. A value of 0.00393 Ω / °C can be used.

[0042] It should be noted that the training process of the physical information neural network includes: constructing a training dataset based on historical power distribution equipment data and corresponding fault labels; defining the neural network structure, including an input layer (receiving standardized equipment data), multiple hidden layers (using the ReLU activation function), and an output layer (using the Sigmoid activation function to output the fault probability); optimizing the network parameters through the backpropagation algorithm to minimize the total loss function, with the number of training iterations ranging from 1000 to 5000.

[0043] Step 2, Anomaly Score Data Analysis: Based on the preliminary fault probability assessment data, perform anomaly scoring analysis to obtain the original anomaly score and generate anomaly score data.

[0044] In a specific example, the step of performing anomaly scoring analysis based on the preliminary fault probability assessment data to obtain an original anomaly score and generating anomaly score data includes: performing data fusion processing based on the preliminary fault probability assessment data and pre-acquired power distribution equipment data to obtain multimodal fusion data; inputting the multimodal fusion data into a pre-trained deep learning model for feature extraction and anomaly detection calculation to obtain an original anomaly score; and performing post-processing optimization on the original anomaly score to generate anomaly score data.

[0045] It should be noted that, based on the preliminary fault probability assessment data and the pre-acquired power distribution equipment data, data fusion processing is performed to obtain multimodal fused data. This includes: firstly, data preprocessing is performed to clean and normalize the data to eliminate dimensional differences and noise effects, generating standardized preliminary fault probability data and standardized multimodal data; then, a weighted fusion algorithm is applied to integrate the standardized data, balancing the contributions of each mode through weights, and dynamically adjusting the weights to adapt to the real-time operating status, for example, increasing the weight of environmental data under high-temperature conditions to enhance the representativeness and anomaly detection capability of the fused data; simultaneously, an autoregressive moving average model is used to predict the system state, providing data for weight adjustment; finally, the weighted fusion results are post-processed, including data smoothing and format conversion, to generate the final multimodal fused data.

[0046] It should be noted that the multimodal fusion data is input into a pre-trained deep learning model for feature extraction and anomaly detection calculation to obtain the original anomaly score. This includes: the deep learning model is a multimodal anomaly detection model based on a self-attention mechanism, which includes a feature extraction function and an anomaly scoring function; wherein the feature extraction function is used to extract deep features from the multimodal fusion data, and the anomaly scoring function is used to calculate the anomaly score.

[0047] Furthermore, through the formula of the feature extraction function The extracted deep feature matrix is ​​obtained. ,in Represented as the input matrix of multimodal fusion data, Represented as a weight matrix, It is represented as a bias vector. It is represented as a rectified linear unit activation function; the deep feature matrix captures nonlinear patterns in multimodal data for subsequent anomaly detection.

[0048] Furthermore, through the anomaly scoring function formula Obtain the original anomaly score ,in This is represented by the number corresponding to the feature. , Represented as the total number of features. Represented as the first The deep feature matrix vector of each feature. This represents the mean vector of the deep feature matrix; the original anomaly score quantifies the degree to which features deviate from the normal pattern, with a larger value indicating a higher probability of anomaly.

[0049] It should be noted that the data fusion processing uses a weighted fusion algorithm, which is expressed by the formula... Completed, among which Represented as multimodal fusion data, Represented as the number corresponding to each mode, , Represented as the total number of modes, Indicates the first Weighting factors corresponding to each mode Indicates the first A data matrix with multiple modalities.

[0050] It should be noted that the modalities are represented as static data, dynamic operational data, status monitoring data, and environmental data. The weight factors corresponding to each modality are obtained through factor analysis. First, the information of each modality is condensed, and then the variance explained after rotation is obtained. The weights are obtained by dividing the cumulative variance explained.

[0051] It should be noted that factor analysis is a well-known technique. It is a multivariate statistical analysis method that starts by studying the internal dependencies of variables and reduces some variables with complex relationships to a few comprehensive factors. Information condensation is expressed as the calculation of the median. The variance explained rate is the amount of information extracted by the factors. Variance explained rate = eigenvalues ​​ / total number of analysis terms. The rotated variance explained rate is expressed as the variance explained by the factors after maximum variance rotation.

[0052] Furthermore, the physical essence of data fusion processing is to perform matrix operations through embedded processors to achieve multi-source data alignment and dimensionality reduction, thereby reducing noise and enhancing the significance of anomalous signals.

[0053] It should be noted that post-processing optimization is performed on the original anomaly scores to generate anomaly score data. This includes: obtaining the original anomaly scores and extracting the minimum and maximum anomaly scores from the training set; calculating the normalized anomaly score, which is equal to the difference between the original anomaly score and the minimum anomaly score, divided by the difference between the maximum and minimum anomaly scores. This scales the anomaly score values ​​to the range of 0 to 1, with early fault symptoms corresponding to anomaly scores greater than 0.7.

[0054] Furthermore, post-processing optimization achieves real-time score adjustment through software modules and dynamically updates thresholds based on historical data to improve detection accuracy.

[0055] This application uses a physical information neural network to generate preliminary fault probability assessment data. It integrates physical models and data-driven methods, reducing false alarm rates by using probability output instead of point estimation, and provides uncertainty quantification, making the prediction results more reliable. This allows for early identification of potential faults and reduces unexpected downtime. Based on the preliminary fault probability assessment data, anomaly scoring analysis is performed, deriving quantitative anomaly scores from the probability assessment. This enables fine-grained anomaly detection, facilitating prioritization and early warning, enhancing the system's sensitivity to abnormal states, and improving the comprehensiveness and response speed of detection.

[0056] Step 3: Risk propagation graph data generation: Based on the anomaly score data and the pre-acquired device network topology data, spatial and temporal dependencies are modeled using a spatiotemporal graph neural network to generate risk propagation graph data.

[0057] It should be noted that the device network topology data is represented as the connection relationship data between devices, including device identifier, connection type and distance weight; the spatiotemporal graph data includes node feature matrix, adjacency matrix and time series data.

[0058] In a specific example, the step of modeling spatial and temporal dependencies based on the anomaly score data and pre-acquired device network topology data using a spatiotemporal graph neural network to generate risk propagation graph data includes: performing graph structure construction processing based on the anomaly score data and pre-acquired device network topology data to generate spatiotemporal graph data; inputting the spatiotemporal graph data into a pre-trained spatiotemporal graph neural network model, processing spatial dependencies through a spatial graph convolutional layer to obtain a spatial feature matrix; further processing temporal dependencies through a temporal convolutional layer to obtain a spatiotemporal fusion feature matrix; and then performing output processing to calculate the mutual influence risk between devices using a risk scoring function to generate risk propagation graph data.

[0059] It should be noted that, based on the abnormal score data and the pre-acquired device network topology data, graph structure construction processing is performed to generate spatiotemporal graph data. The specific process is as follows: The graph structure construction processing is performed by an embedded graph processing module. Based on the device identifier, the abnormal score data is mapped to node features, and then the spatial relationship between devices is calculated through the adjacency matrix generation algorithm to generate an adjacency matrix. At the same time, time series data, including the time change data of abnormal scores, is integrated to form spatiotemporal graph data.

[0060] Furthermore, the formula for the adjacency matrix generation algorithm is: ,in Represents the adjacency matrix of the nth element. The device and the first The connection status of a device is represented by 1, where 0 indicates a connection and 1 indicates a non-connection. The connection status is determined based on the connection type and distance weight in the device network topology data. If the distance weight is less than a preset threshold, the device is set to connected.

[0061] It should be noted that the spatiotemporal graph data is input into a pre-trained spatiotemporal graph neural network model, and spatial dependencies are processed through spatial graph convolutional layers to obtain the spatial feature matrix. The specific process is as follows: the spatial graph convolutional layer uses a graph convolutional network algorithm, and the formula is... ,in This is represented by the node number corresponding to the node feature matrix. Represented as the first The node feature matrix of the layer Represented as the first The output feature matrix of the layer, This is represented as an adjacency matrix with added self-connections. Represented as an identity matrix, Represented as The degree matrix, Represented as the first The trainable weight matrix of the layer, This is represented as an activation function.

[0062] Furthermore, the node feature matrix represents the intermediate features of devices in the risk propagation process, such as the diffusion state of abnormal score data; the trainable weight matrix represents the neural network parameters, which learn the weights for risk propagation through training; the degree matrix is ​​used for normalization to prevent gradient explosion; the spatial graph convolutional layer captures the spatial risk propagation between devices by aggregating neighbor node information; this process is accelerated by GPU computation to achieve efficient spatial feature extraction.

[0063] It should be noted that the temporal convolutional layer is used to process temporal dependencies to obtain a spatiotemporal fusion feature matrix; the temporal convolutional layer captures the changing patterns of abnormal scores over time through sliding window convolution, thereby modeling the temporal propagation of risk; the physical essence of the temporal convolutional layer is to process the historical trend of device risk, such as the fluctuation of abnormal scores, and extract key patterns in the time series through convolution operations to enhance the model's adaptability to dynamic risks.

[0064] In a specific example, the output processing, which calculates the mutual influence risk between devices using a risk scoring function to generate risk propagation map data, includes: A1, using a risk scoring function... The result is the The device and the first Risk propagation score of each device , and Represent the spatiotemporal fusion feature matrix, respectively. The device and the first Feature vectors of each device The function is represented as compressing the output to a probability range.

[0065] A2. Based on the risk propagation score, post-processing optimization is performed, including normalization and threshold filtering, to generate the final risk propagation map data, which includes the risk score matrix corresponding to all devices.

[0066] It should be noted that the risk propagation graph data includes a risk score matrix corresponding to all devices, representing the risk propagation relationship between devices; the risk propagation score value is between 0 and 1, and the larger the value, the higher the risk of mutual influence; the physical meaning of the risk scoring function is to quantify the contagion of risk between devices, such as the degree of influence of a device with a high anomaly score on its neighboring devices; this processing is implemented through matrix multiplication, utilizing hardware accelerators to improve computational efficiency, and outputting a visual risk graph for subsequent decision-making.

[0067] It should be noted that the post-processing optimization first normalizes the risk score matrix and uses min-max scaling to make the risk scores range between 0 and 1; then threshold filtering is applied to remove connections with scores below a preset threshold (such as 0.3) to reduce noise.

[0068] Step 4: Maintenance strategy data generation: Based on the risk propagation map data and the pre-acquired maintenance resource data, maintenance decision calculations are performed to obtain the original maintenance action probability distribution and priority score, so as to generate the optimal maintenance strategy data.

[0069] It should be noted that maintenance resource data includes resource information such as the number of maintenance personnel, tool availability, and time constraints.

[0070] In a specific example, the step of calculating maintenance decisions based on the risk propagation graph data and pre-acquired maintenance resource data to obtain the original maintenance action probability distribution and priority score, in order to generate optimal maintenance strategy data, includes: performing state vector construction processing based on the risk propagation graph data and pre-acquired maintenance resource data to obtain a state representation vector; then inputting the state representation vector into a pre-trained deep reinforcement learning model, performing maintenance decision calculation through a policy network to obtain the original maintenance action probability distribution and priority score; and performing decision optimization processing on the original maintenance action probability distribution and priority score to generate optimal maintenance strategy data.

[0071] It should be noted that, based on the risk propagation map data and the pre-acquired maintenance resource data, a state vector construction process is performed to obtain a state representation vector. This process includes: the state vector construction process, based on the risk score matrix and maintenance resource data vector in the risk propagation map data, generates a state representation vector through vector concatenation and normalization operations; wherein the formula is used... Derive the state representation vector ,in Represented as a risk score matrix, This is represented as a matrix flattening operation. Represented as a maintenance resource vector, This is represented as a vector concatenation operation.

[0072] Furthermore, the risk score matrix represents the risk propagation relationship between devices; the maintenance resource vector represents available maintenance resources; and a normalization operation is applied to the state representation vector to scale each eigenvalue to... The range is adjusted to improve the stability of model inputs.

[0073] It should be noted that the state representation vector is input into a pre-trained deep reinforcement learning model, and maintenance decisions are calculated through a policy network to obtain the original maintenance action probability distribution and priority score. This includes: the pre-trained deep reinforcement learning model employs the policy gradient method, where the policy network formula is... , indicating that in the state representation vector Select maintenance action when the corresponding status is reached. The probability of, where Represented as network parameters, The state is represented as a maintenance action; the policy network takes a state representation vector as input and outputs a maintenance action probability distribution and priority score.

[0074] Furthermore, maintenance actions include various types, such as repair, replacement, and inspection. Therefore, the maintenance action probability distribution is a probability vector, with each element corresponding to the probability of an action. The priority score is a scalar representing the urgency of the maintenance. The policy network structure includes an input layer, multiple hidden layers (using the ReLU activation function), and an output layer. The output layer uses the Softmax activation function to transform the output into a probability distribution, and a linear activation function to output the priority score.

[0075] In a specific example, the decision optimization process of the original maintenance action probability distribution and priority score to generate optimal maintenance strategy data includes: the decision optimization process selects the maintenance action with the highest probability based on the original maintenance action probability distribution as the final action; and the process discretizes the priority based on the priority score to determine the maintenance priority.

[0076] It should be noted that discretization is achieved through preset thresholds, such as mapping priority scores to high, medium, and low priority levels; decision optimization is achieved by the embedded processor performing maximum value selection and comparison operations to generate executable maintenance instructions, ensuring the real-time performance and feasibility of the maintenance strategy.

[0077] This application uses a spatiotemporal graph neural network to model spatial and temporal dependencies, generating risk propagation graph data to capture dynamic risk propagation in the device network. This improves the real-time performance and accuracy of risk assessment, helps identify risk hotspots and propagation paths, thereby optimizing resource allocation and reducing failures. Based on risk propagation graph data and maintenance resource data, maintenance decision calculations are performed, integrating multi-source data to generate optimal maintenance strategies. This improves maintenance efficiency and quality, reduces maintenance costs and time, and emphasizes probabilistic optimization and resource constraints.

[0078] Step 5: Dynamic Risk Profile Data Generation: Attribution analysis is performed based on the optimal maintenance strategy data to generate dynamic risk profile data.

[0079] In a specific example, the attribution analysis based on optimal maintenance strategy data to generate dynamic risk profile data, displaying the main risk factors and their contributions, includes: performing data integration processing based on the optimal maintenance strategy data and the data from steps one, two, and three to obtain attribution analysis input data; inputting the attribution analysis input data into a pre-trained interpretable artificial intelligence model, performing feature importance analysis through the attribution calculation layer to obtain raw contribution scores, performing normalization processing to obtain standardized contribution scores; and generating dynamic risk profile data based on the standardized contribution scores, wherein the dynamic risk profile data includes the main risk factors and their corresponding contributions.

[0080] It should be noted that, based on the optimal maintenance strategy data and the intermediate data from steps one to three, data integration processing is performed to obtain the attribution analysis input data. The specific process is as follows: Data integration processing generates an attribution analysis input data matrix based on the optimal maintenance strategy data, preliminary fault probability assessment data, anomaly score data, and risk propagation map data; firstly, through a vector concatenation function... The preliminary failure probability assessment data vector, anomaly score data vector, risk propagation map data vector, and optimal maintenance strategy data vector are concatenated, and then standardized using min-max scaling to ensure that all feature values ​​fall within the specified range. Intervals are used to eliminate differences in dimensions.

[0081] It should be noted that the attribution analysis input data is fed into a pre-trained interpretable artificial intelligence model, and feature importance analysis is performed through the attribution calculation layer to obtain the raw contribution score. Normalization is then performed to obtain the standardized contribution score, including: the interpretable artificial intelligence model is an attribution model based on the attention mechanism, and its attribution calculation layer uses attention weights to calculate feature importance. Firstly, it can be explained that the AI ​​model multiplies the input data matrix with the trainable weight matrix to generate a query matrix, a key matrix, and a value matrix. The query matrix represents a transformed form of the input features, used to capture dependencies between features; the key matrix represents feature keys, used to calculate similarity; and the value matrix represents feature values, used for weighted summation. Secondly, it can be explained that the AI ​​model calculates the attention score matrix by performing a dot product operation on the transpose of the query matrix and the key matrix, then scaling by the square root of the attention head dimension to prevent gradient explosion, and finally applying the softmax function to convert the score into a probability distribution. Finally, it can be explained that the AI ​​model outputs an attention weight matrix obtained by multiplying the attention score matrix with the value matrix. The elements in the attention score matrix quantify the importance weights between features, representing the relative contribution of each feature to the overall risk in risk analysis, thus obtaining the original contribution score. A larger value indicates a greater impact of the feature on the risk.

[0082] Furthermore, the ratio of the original contribution score minus the minimum original contribution score to the original contribution score minus the minimum original contribution score is used as the standardized contribution score.

[0083] It should be noted that dynamic risk profile data is generated based on standardized contribution. This dynamic risk profile data includes major risk factors and their corresponding contribution. The process involves: selecting features with a standardized contribution greater than a preset contribution threshold (e.g., 0.7) as major risk factors through threshold filtering and sorting operations based on standardized contribution, and recording their contribution values ​​to form a list of key-value pairs. This list is then encapsulated into a JSON data structure, including timestamps and device identifiers, to generate dynamic risk profile data.

[0084] Furthermore, the main risk factors and their contributions are displayed in real time through a graphical user interface, such as using bar charts or heatmaps, allowing users to intuitively understand the risk distribution.

[0085] The dynamic risk profiling in this application supports real-time monitoring and decision-making, and reveals root causes through attribution analysis, thereby enhancing the transparency and interpretability of the system and assisting operations and maintenance personnel in responding quickly.

[0086] Please see Figure 2As shown, in its second aspect, this application provides a system for a comprehensive quantitative assessment method of power distribution equipment failure probability.

[0087] The system 100 of the comprehensive quantitative assessment method for the failure probability of power distribution equipment described in this invention can be installed in an electronic device. Depending on the functions implemented, the system 100 may include a preliminary failure probability analysis module 101, an anomaly score data analysis module 102, a risk propagation map data generation module 103, a maintenance strategy data generation module 104, and a dynamic risk profile data generation module. The modules described in this invention can also be referred to as units, which are a series of computer program segments that can be executed by the processor of an electronic device and perform a fixed function, stored in the memory of the electronic device.

[0088] In this embodiment, the functions of each module / unit are as follows:

[0089] The preliminary fault probability analysis module, based on pre-acquired power distribution equipment data, processes the data through a physical information neural network to generate preliminary fault probability assessment data.

[0090] The anomaly score data analysis module performs anomaly scoring analysis based on the preliminary fault probability assessment data to obtain the original anomaly score and generate anomaly score data.

[0091] The risk propagation graph data generation module, based on the anomaly score data and the pre-acquired device network topology data, uses a spatiotemporal graph neural network to model spatial and temporal dependencies in order to generate risk propagation graph data;

[0092] The maintenance strategy data generation module, based on the risk propagation map data and the pre-acquired maintenance resource data, performs maintenance decision calculations to obtain the original maintenance action probability distribution and priority score, so as to generate the optimal maintenance strategy data;

[0093] The dynamic risk profile data generation module performs attribution analysis based on optimal maintenance strategy data to generate dynamic risk profile data.

[0094] In the several embodiments provided by this invention, it should be understood that the disclosed methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.

[0095] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0096] This application provides a comprehensive quantitative assessment method and system for the probability of power distribution equipment failure. By integrating physical information neural networks, anomaly scoring analysis, spatiotemporal graph neural networks, and maintenance decision calculations, it achieves intelligent management of power distribution equipment failure prediction and maintenance strategies, improves the accuracy and reliability of failure prediction, reduces false alarm rates through probability output, and supports uncertainty quantification; enhances the sensitivity of anomaly detection, facilitating early warning and prioritization, and improving the system's response speed to abnormal states; comprehensively models spatial and temporal dependencies, optimizes risk assessment and resource allocation, and reduces the risk of cascading failures; generates data-driven maintenance strategies, improving maintenance efficiency and quality, and reducing costs and downtime; and generates dynamic risk profiles through attribution analysis, supporting real-time monitoring and decision-making, and enhancing system transparency.

[0097] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0098] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0099] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0100] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A comprehensive quantitative assessment method for the failure probability of power distribution equipment, characterized in that, include: Step 1: Preliminary Fault Probability Analysis: Based on the pre-acquired power distribution equipment data, the data is processed through a physical information neural network to generate preliminary fault probability assessment data. Step 2, Anomaly Score Data Analysis: Based on the preliminary fault probability assessment data, perform anomaly scoring analysis to obtain the original anomaly score and generate anomaly score data; Step 3: Risk propagation graph data generation: Based on the anomaly score data and the pre-acquired device network topology data, spatial and temporal dependencies are modeled using a spatiotemporal graph neural network to generate risk propagation graph data; Step 4: Maintenance strategy data generation: Based on the risk propagation map data and the pre-acquired maintenance resource data, maintenance decision calculations are performed to obtain the original maintenance action probability distribution and priority score, so as to generate the optimal maintenance strategy data; Step 5: Dynamic Risk Profile Data Generation: Attribution analysis is performed based on the optimal maintenance strategy data to generate dynamic risk profile data.

2. The method for comprehensive quantitative assessment of power distribution equipment failure probability according to claim 1, characterized in that, The pre-acquired power distribution equipment data includes static data, dynamic operation data, status monitoring data, and environmental data.

3. The method for comprehensive quantitative assessment of power distribution equipment failure probability according to claim 1, characterized in that, The pre-acquired power distribution equipment data is processed through a physical information neural network to generate preliminary fault probability assessment data, including: Based on the pre-acquired power distribution equipment data, data cleaning and normalization are performed to obtain standardized power distribution equipment data. The standardized power distribution equipment data is then input into a pre-trained physical information neural network. Through the forward propagation calculation of the physical information neural network, preliminary fault probability assessment data is generated. Physical mechanism constraints are incorporated into the physical information neural network during training.

4. The method for comprehensive quantitative assessment of power distribution equipment failure probability according to claim 1, characterized in that, The process of performing anomaly scoring analysis based on the preliminary fault probability assessment data to obtain raw anomaly scores and generating anomaly score data includes: Based on the preliminary fault probability assessment data and the pre-acquired power distribution equipment data, data fusion processing is performed to obtain multimodal fusion data; the multimodal fusion data is input into a pre-trained deep learning model for feature extraction and anomaly detection calculation to obtain the original anomaly score; the original anomaly score is post-processed and optimized to generate anomaly score data.

5. The method for comprehensive quantitative assessment of power distribution equipment failure probability according to claim 1, characterized in that, The process of generating risk propagation graph data by modeling spatial and temporal dependencies using a spatiotemporal graph neural network based on the anomaly score data and pre-acquired device network topology data includes: Based on the aforementioned abnormal score data and pre-acquired device network topology data, graph structure construction processing is performed to generate spatiotemporal graph data. The spatiotemporal graph data is then input into a pre-trained spatiotemporal graph neural network model, where spatial dependencies are processed through spatial graph convolutional layers to obtain a spatial feature matrix. Temporal dependencies are then processed through temporal convolutional layers to obtain a spatiotemporal fusion feature matrix. Finally, output processing is performed, and the mutual influence risk between devices is calculated using a risk scoring function to generate risk propagation graph data.

6. The method for comprehensive quantitative evaluation of power distribution equipment failure probability according to claim 5, characterized in that, The output processing involves calculating the mutual influence risk between devices using a risk scoring function to generate risk propagation map data, including: A1. Through risk scoring function The result is the The device and the first Risk propagation score of each device , and Represent the spatiotemporal fusion feature matrix, respectively. The device and the first Feature vectors of each device The function represents compressing the output to a probability range; A2. Based on the risk propagation score, post-processing optimization is performed, including normalization and threshold filtering, to generate the final risk propagation map data, which includes the risk score matrix corresponding to all devices.

7. The method for comprehensive quantitative assessment of power distribution equipment failure probability according to claim 1, characterized in that, Based on the risk propagation map data and pre-acquired maintenance resource data, maintenance decision calculations are performed to obtain the original maintenance action probability distribution and priority score, in order to generate optimal maintenance strategy data, including: Based on the risk propagation graph data and the pre-acquired maintenance resource data, state vector construction is performed to obtain state representation vectors. Then, the state representation vectors are input into a pre-trained deep reinforcement learning model, and maintenance decision calculation is performed through a policy network to obtain the original maintenance action probability distribution and priority score. The original maintenance action probability distribution and priority score are then subjected to decision optimization processing to generate optimal maintenance strategy data.

8. The method for comprehensive quantitative assessment of power distribution equipment failure probability according to claim 7, characterized in that, The step of performing decision optimization processing on the probability distribution and priority scores of the original maintenance actions to generate optimal maintenance strategy data includes: The decision optimization process selects the maintenance action with the highest probability based on the probability distribution of the original maintenance actions as the final action; and performs discretization based on the priority score to determine the maintenance priority.

9. The method for comprehensive quantitative assessment of power distribution equipment failure probability according to claim 1, characterized in that, The attribution analysis based on optimal maintenance strategy data generates dynamic risk profile data, displaying the main risk factors and their contribution, including: Based on the optimal maintenance strategy data and the data from steps one, two, and three, data integration processing is performed to obtain attribution analysis input data. The attribution analysis input data is then input into a pre-trained interpretable artificial intelligence model, and feature importance analysis is performed through the attribution calculation layer to obtain the raw contribution score. Normalization processing is then performed to obtain the standardized contribution score. Based on the standardized contribution score, dynamic risk profile data is generated, which includes the main risk factors and their corresponding contributions.

10. A system for implementing the comprehensive quantitative assessment method for the failure probability of power distribution equipment as described in any one of claims 1-9, characterized in that, include: The preliminary fault probability analysis module, based on pre-acquired power distribution equipment data, processes the data through a physical information neural network to generate preliminary fault probability assessment data. The anomaly score data analysis module performs anomaly scoring analysis based on the preliminary fault probability assessment data to obtain the original anomaly score and generate anomaly score data. The risk propagation graph data generation module, based on the anomaly score data and the pre-acquired device network topology data, uses a spatiotemporal graph neural network to model spatial and temporal dependencies in order to generate risk propagation graph data; The maintenance strategy data generation module, based on the risk propagation map data and the pre-acquired maintenance resource data, performs maintenance decision calculations to obtain the original maintenance action probability distribution and priority score, so as to generate the optimal maintenance strategy data; The dynamic risk profile data generation module performs attribution analysis based on optimal maintenance strategy data to generate dynamic risk profile data.

Images