A power distribution network fault intelligent diagnosis method, system, device and medium

Abstract

The present application relates to a kind of power distribution network fault intelligent diagnosis method, system, equipment and medium, method includes: the quasi-steady state of the power distribution network measured data, dynamic synchronous phasor data and meteorological data obtained are space-time alignment, construct multi-source tensor;The information transmission direction and intensity between node characteristic sequence are calculated and significant screening is carried out, generate dynamic causal diagram;Introduce structural equation model to effect quantization of node causal relationship, solve structural equation and obtain causal structure function and calculate average causal effect, generate causal effect matrix;Based on dynamic causal diagram and causal effect matrix, construct causal enhancement graph neural network, the causal effect matrix is embedded as causal weight in network propagation process and is trained, generates diagnostic model;Based on causal structure function, the mixed factors in the sample to be diagnosed are counterfactual intervention processing and generate counterfactual sample set, determine fusion weight by combining causal effect matrix to weighted fusion of diagnostic probability, generate final diagnosis result.

Description

Technical Field

[0001] This invention relates to the field of power system automation technology, specifically to a method, system, equipment, and medium for intelligent fault diagnosis in power distribution networks. Background Technology

[0002] With the deepening of the construction of new power systems, the distribution network, as a crucial link connecting the power supply side and the user side, has seen its operational safety and intelligence level become key indicators for measuring the stability and resilience of modern power systems. In recent years, the large-scale integration of distributed renewable energy sources, the dynamic changes of various load types, and the dynamic reconfiguration of network topology have resulted in distribution networks exhibiting high-dimensionality, strong randomness, and complex dynamic evolution characteristics. Simultaneously, the widespread adoption of intelligent sensing technology has driven the extensive deployment of multi-source measurement terminals such as SCADA systems, PMU measurement devices, and meteorological monitoring terminals, leading to distribution network operational data exhibiting high-dimensionality, heterogeneity, and asynchronous characteristics. How to effectively integrate multiple types of data and accurately identify fault characteristics from complex operating states has become a research hotspot and challenge in the field of distribution network fault diagnosis.

[0003] Currently, while deep learning-based fault diagnosis methods for distribution networks are widely used, they are essentially "black box" pattern recognition based on statistical correlation. These methods suffer from a fundamental flaw: when the input data contains confounding factors unrelated to the fault (such as weather disturbances, load fluctuations, and measurement noise), deep learning models are prone to learning spurious correlations between variables rather than genuine causal mechanisms. For example, the model might misjudge the co-occurrence of "thunderstorms" and "line faults" as a causal relationship, leading to a sharp drop in diagnostic accuracy when the operating environment changes. Furthermore, existing models often focus on structural correlation modeling, lacking a deep characterization of the causal mechanisms during the dynamic evolution of the distribution network, resulting in problems such as weak generalization ability and difficulty in interpreting diagnostic results in practical applications. Summary of the Invention

[0004] Based on this, the purpose of this invention is to provide a method, system, device, and medium for intelligent diagnosis of power distribution network faults that can remove confounding factors from observation data and learn the true causal mechanism.

[0005] The objective of this invention is achieved through the following solution:

[0006] In a first aspect, the present invention provides an intelligent fault diagnosis method for power distribution networks, comprising the following steps:

[0007] S1: The acquired quasi-steady-state measurement data of the distribution network, dynamic synchronous phasor data and meteorological data are processed for spatiotemporal alignment using the dynamic time warping algorithm to construct a multi-source tensor containing three dimensions: time, node and feature.

[0008] S2: Based on multi-source tensors, the information transmission direction and intensity between feature sequences of any two nodes are calculated using the transfer entropy algorithm. The calculated information transmission intensity is then filtered for significance, and causal relationship edges with statistical significance are selected to generate a dynamic causal graph.

[0009] S3: Based on dynamic causal graphs and multi-source tensors, a structural equation model is introduced to quantify the causal relationships of each node. The causal structure function of each node is obtained by solving the structural equation, and the average causal effect between nodes is calculated based on the causal structure function to generate a causal effect matrix.

[0010] S4: Construct a causal enhanced graph neural network based on dynamic causal graph and causal effect matrix. Use dynamic causal graph as network graph structure and embed causal effect matrix as causal weights for message passing into the network propagation process. Use historical data of multi-source tensors with fault type and location as supervision labels for training to generate a diagnostic model that integrates causal prior knowledge.

[0011] S5: Based on the causal structure function, counterfactual intervention is performed on the confounding factors in the samples to be diagnosed obtained from the real-time data stream to generate a counterfactual sample set. The samples to be diagnosed and the counterfactual sample set are respectively input into the diagnostic model for probabilistic diagnosis. The fusion weight is determined by combining the effect intensity of the confounding factors in the causal effect matrix to perform weighted fusion of the diagnostic probabilities and generate the final diagnostic result after removing the interference of confounding factors.

[0012] In one embodiment, S1 of the intelligent fault diagnosis method for power distribution networks provided by the present invention specifically includes the following steps:

[0013] S11: Acquire quasi-steady-state measurement data of each node in the distribution network, dynamic synchronous phasor data of each PMU deployment location, and meteorological data of each meteorological monitoring station. Perform missing value imputation and outlier removal on various types of data to generate a cleaned multi-source dataset.

[0014] S12: Normalize the various types of data in the cleaned multi-source dataset, map features of different dimensions to a unified numerical range, and generate a standardized feature sequence.

[0015] S13: Perform time axis alignment processing on the feature sequence, calculate the optimal alignment path between data sequences with different sampling frequencies, eliminate the phase difference caused by communication delay through path mapping, and generate time-aligned multi-source data;

[0016] S14: Organize the multi-source data into a three-dimensional multi-source tensor by tensing according to three dimensions: timestamp, node number, and feature type.

[0017] In one embodiment, S2 of the intelligent fault diagnosis method for distribution networks provided by the present invention specifically includes the following steps:

[0018] S21: Perform sliding window partitioning on the voltage and current feature sequences of each node in the multi-source tensor, calculate the mutual information value between any two node feature sequences within each window, and construct an initial adjacency matrix based on the mutual information value;

[0019] S22: Based on the initial adjacency matrix, the direction of information transmission between each pair of nodes is calculated and processed. By comparing the transmission entropy values ​​in different directions, the dominant direction of information flow is determined, and a directed graph structure with directional weights is generated.

[0020] S23: Perform significance testing on each directed edge in the directed graph structure, filter out weakly related edges with a propagation entropy value lower than the preset confidence level, retain causal relationship edges with statistical significance, and generate a dynamic causal graph.

[0021] In one embodiment, S3 of the intelligent fault diagnosis method for power distribution networks provided by the present invention specifically includes the following steps:

[0022] S31: Based on the set of parent nodes of each node in the dynamic causal graph, structural equation modeling is performed on the mapping relationship between the state features of each node and the features of the parent node. Basis function expansion and fitting are performed on the mapping relationship between the state features and the features of the parent node. The basis function weight coefficients are estimated by the least squares method to generate the initial structure function of each node.

[0023] S32: Perform additivity decomposition on the initial structure function to separate the contributions of exogenous and endogenous variables. Iteratively optimize the structure function parameters by minimizing the prediction error to generate the causal structure function for each node.

[0024] S33: Based on the causal structure function, a forward propagation calculation is performed after applying a unit impulse perturbation to each node. After applying a unit impulse perturbation to the causal structure function of each node, a forward inference process is performed. By comparing the influence of the changes in the characteristics of the parent node set before and after the impulse is applied on the characteristics of the child node, a causal effect matrix is ​​generated by traversing all node pairs.

[0025] In one embodiment, the formula for calculating the causal effect matrix of the intelligent fault diagnosis method for distribution networks provided by the present invention is as follows:

[0026]

[0027] in, Let be the strength of the modified causal effect of node j on node i. Let i be the causal structure function of node i. Let be the set of parent nodes of node i at time t. Let represent the state characteristics of node j at time t. For time-varying decay factor control parameters, Let j be the information propagation delay from node j to node i. Let the oscillation characteristic period between node j and node i be . Let t be the phase difference between node j and node i at time t, and T be the total time length.

[0028] In one embodiment, S4 of the intelligent fault diagnosis method for power distribution networks provided by the present invention specifically includes the following steps:

[0029] S41: Based on the topological structure of the dynamic causal graph, construct the adjacency matrix of the graph neural network, embed the elements in the causal effect matrix as the initial causal weights for message passing into the corresponding positions of the adjacency matrix, perform causal effect normalization on the weight values ​​of each edge in the adjacency matrix, and construct a graph neural network framework that integrates causal weights.

[0030] S42: Perform multi-head attention calculation on the hidden state of each node in the graph neural network framework, and concatenate and fuse the node's own features, neighbor node features and corresponding causal effect weights. Adaptively adjust the contribution of different causal paths through learnable attention parameters to generate causally enhanced node representations.

[0031] S43: The graph neural network framework is trained end-to-end based on historical data from multi-source tensors. Fault type labels and fault location labels are used as supervision signals for backpropagation optimization. The network parameters are iteratively updated through gradient descent algorithm until the loss function converges, generating a diagnostic model that integrates causal prior knowledge.

[0032] In one embodiment, S5 of the intelligent fault diagnosis method for power distribution networks provided by the present invention specifically includes the following steps:

[0033] S51: Perform confounding factor identification processing on the samples to be diagnosed obtained from the real-time data stream, mark meteorological features and load fluctuation features as confounding variables to be intervened, and extract the state features of each node in the samples to be diagnosed as feature benchmark values ​​for counterfactual reasoning.

[0034] S52: Based on the causal structure function, different combinations of intervention values ​​are applied to the confounding variables in the feature baseline value in sequence, while the non-confounding variables are kept unchanged. The forward inference calculation of the state features of each node after intervention is performed through the causal structure function to generate multiple sets of counterfactual state features after intervention.

[0035] S53: Reorganize the counterfactual state features according to the dimensional organization of multi-source tensors to construct a counterfactual sample set containing multiple counterfactual samples;

[0036] S54: Input the sample to be diagnosed into the diagnostic model for forward propagation calculation, and map it layer by layer through graph convolutional layers and fully connected layers to output the original diagnostic probability vector of each node belonging to various types of faults;

[0037] S55: Input each sample in the counterfactual sample set into the diagnostic model in sequence for parallel computation, and obtain the corresponding counterfactual diagnostic probability vector set respectively;

[0038] S56: Based on the mean effect intensity of confounding variables on the voltage and current characteristics of each node in the causal effect matrix, calculate the fusion weight coefficient, and perform weighted averaging on the original diagnostic probability vector and the counterfactual diagnostic probability vector set according to the fusion weight coefficient to generate the final diagnostic result after removing the interference of confounding factors.

[0039] Secondly, the present invention provides an intelligent fault diagnosis system for power distribution networks, which is configured with the following modules:

[0040] The multi-source data spatiotemporal alignment module is used to perform spatiotemporal alignment processing on the acquired quasi-steady-state measurement data of the distribution network, dynamic synchronous phasor data and meteorological data using a dynamic time warping algorithm, and to construct a multi-source tensor containing three dimensions: time, node and feature.

[0041] The dynamic causal graph construction module is used to calculate the information transmission direction and intensity between feature sequences of any two nodes based on multi-source tensors using the transfer entropy algorithm, and to perform significance filtering on the calculated information transmission intensity to select causal relationship edges with statistical significance, thereby generating a dynamic causal graph.

[0042] The causal effect quantification module is used to quantify the causal relationships of each node based on the dynamic causal graph and multi-source tensor, and introduces the structural equation model. It obtains the causal structure function of each node by solving the structural equation, and calculates the average causal effect between nodes based on the causal structure function to generate the causal effect matrix.

[0043] The Causal Augmented Graph Network Training Module is used to construct a causal augmented graph neural network based on a dynamic causal graph and a causal effect matrix. The dynamic causal graph is used as the network graph structure and the causal effect matrix is ​​embedded as the causal weights for message passing in the network propagation process. Historical data from multi-source tensors are used for training with fault type and location as supervision labels to generate a diagnostic model that integrates causal prior knowledge.

[0044] The counterfactual diagnostic fusion module is used to perform counterfactual intervention processing on the confounding factors in the samples to be diagnosed obtained from the real-time data stream based on the causal structure function, generate a counterfactual sample set, input the samples to be diagnosed and the counterfactual sample set into the diagnostic model for probabilistic diagnosis, and combine the effect strength of the confounding factors in the causal effect matrix to determine the fusion weight to perform weighted fusion of the diagnostic probabilities, and generate the final diagnostic result after removing the interference of confounding factors.

[0045] Thirdly, this application provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement any of the above-mentioned intelligent fault diagnosis methods for power distribution networks.

[0046] Fourthly, this application provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements any of the above-mentioned intelligent fault diagnosis methods for power distribution networks.

[0047] In summary, the intelligent fault diagnosis method for distribution networks provided in this application can accurately capture the real information transmission paths between nodes during the operation of the distribution network by constructing a dynamic causal graph using the transfer entropy algorithm, effectively filtering out spurious correlations caused by meteorological interference and load fluctuations. The introduction of a causal effect matrix quantified by structural equation modeling transforms abstract causal strength into quantifiable computational weights, providing a physically interpretable guiding basis for subsequent message transmission in the neural network. Embedding the causal effect matrix as prior knowledge into the message transmission process of the graph neural network guides the model to focus on the features of neighboring nodes that truly have causal influence, avoiding the attention mechanism being misled by noisy data. Based on the counterfactual sample set generated by the causal structure function, the system state under different scenarios can be simulated by intervening in confounding factors, enabling robust verification of diagnostic results under environmental changes. Combining the causal effect strength to determine the fusion weights and weightedly fusing the probabilities of the original diagnosis and counterfactual diagnosis can effectively suppress the risk of misjudgment caused by changes in operating mode or sudden changes in the external environment, thereby achieving the model's adaptive adjustment capability under multiple operating conditions and ultimately improving the credibility and interpretability of the diagnostic results.

[0048] To better understand and implement this invention, the following detailed description is provided in conjunction with the accompanying drawings. Attached Figure Description

[0049] Figure 1 A flowchart illustrating an intelligent fault diagnosis method for a power distribution network provided in this application embodiment;

[0050] Figure 2 A schematic diagram of the process for generating a causal effect matrix provided in an embodiment of this application;

[0051] Figure 3This is a schematic diagram of the structure of a power distribution network fault intelligent diagnosis system provided in another embodiment of this application. Detailed Implementation

[0052] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Preferred embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the invention.

[0053] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0054] In one embodiment, such as Figure 1 As shown, a method for intelligent fault diagnosis in a distribution network is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, or to a system including both a terminal and a server, and is implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0055] S1: The acquired quasi-steady-state measurement data of the distribution network, dynamic synchronous phasor data and meteorological data are processed for spatiotemporal alignment using the dynamic time warping algorithm to construct a multi-source tensor containing three dimensions: time, node and feature.

[0056] Specifically, the system acquires quasi-steady-state measurement data, dynamic synchronization phasor data, and meteorological data during the operation of the distribution network. Quasi-steady-state measurement data is collected through a SCADA system, covering electrical quantities such as voltage, current, active power, and reactive power at each node of the distribution network. Dynamic synchronization phasor data is collected through PMU measurement devices, covering high-frequency dynamic electrical quantities such as synchronization phasors, frequency, and phase angle at each node. Meteorological data is collected through meteorological monitoring terminals deployed in different areas of the distribution network, covering temperature, humidity, wind speed, rainfall, and thunderstorm-related data in the operating area of ​​the distribution network. Due to differences in sampling frequencies between the acquisition terminals for different types of data, and the time offset between acquisition points at different spatial nodes, spatiotemporal misalignment occurs in the multi-source data, making it unsuitable for direct use in subsequent causal analysis and fault diagnosis. The system employs a dynamic time warping algorithm to perform spatiotemporal alignment processing on the three types of data. This algorithm flexibly adjusts the time axis, calculates the similarity between different time series, and achieves time axis matching for data with different sampling frequencies. Simultaneously, based on the spatial distribution of distribution network nodes, it maps the acquired data from different nodes to a unified spatial coordinate system, completing the spatial dimension alignment.

[0057] Furthermore, the system integrates and reconstructs the aligned quasi-steady-state measurement data, dynamic synchronous phasor data, and meteorological data according to three dimensions: time series, distribution network node number, and electrical and meteorological characteristics, to form a multi-source tensor. The time dimension of the multi-source tensor corresponds to the temporal sequence of data acquisition, the node dimension corresponds to all monitoring nodes of the distribution network, and the characteristic dimension corresponds to various electrical and meteorological parameters. Through this multi-source tensor, the system achieves a standardized and unified expression of multi-source heterogeneous data, eliminating data deviations caused by asynchronous sampling and spatiotemporal misalignment.

[0058] S2: Based on multi-source tensors, the information transmission direction and intensity between feature sequences of any two nodes are calculated using the transfer entropy algorithm. The calculated information transmission intensity is then filtered for significance, and causal relationship edges with statistical significance are selected to generate a dynamic causal graph.

[0059] Specifically, the system invokes the transfer entropy algorithm to calculate the characteristic sequences of any two distribution network nodes in the multi-source tensor. The transfer entropy value quantifies the direction and intensity of information transfer between the characteristic sequences of the two nodes. The transfer entropy algorithm is a nonlinear causal correlation measurement method that does not require the assumption of a linear relationship between variables. It can effectively distinguish between correlation and causal relationships between variables, accurately capture the asymmetric information transfer mechanism between nodes, and avoid the limitations of linear analysis methods that cannot adapt to the nonlinear operating characteristics of distribution networks.

[0060] Preferably, the system performs statistical significance screening on the information transmission strength calculation results of all node pairs. Using hypothesis testing methods and setting testing criteria, it eliminates non-significant correlation values ​​generated by irrelevant factors such as random noise and normal load fluctuations, retaining causal relationship edges that pass the significance test and possess real physical meaning. The system uses each monitoring node of the distribution network as a node in the graph and the screened causal relationship edges as connecting edges. Based on the real-time operating status of the distribution network reflected by the multi-source tensor, it dynamically generates a dynamic causal graph. This dynamic causal graph can update the causal relationships between nodes in real time, following changes in the operating status of the distribution network, clearly representing the dynamically evolving causal relationship topology between nodes in the distribution network. Through this dynamic causal graph, the system transforms the causal relationships between distribution network nodes into a visualized and structured topological form.

[0061] S3: Based on dynamic causal graphs and multi-source tensors, a structural equation model is introduced to quantify the causal relationships of each node. The causal structure function of each node is obtained by solving the structural equation, and the average causal effect between nodes is calculated based on the causal structure function to generate a causal effect matrix.

[0062] Specifically, based on the constructed dynamic causal graph and multi-source tensor, the system introduces structural equation modeling to quantify the causal relationships between nodes in the dynamic causal graph. Structural equation modeling can accurately characterize the causal dependencies between variables through a system of equations. The system uses various types of data from the multi-source tensor as observed variables and the topological structure of the dynamic causal graph as constraints to construct a system of structural equations representing the causal dependencies between node variables. Each equation in the system of structural equations corresponds to a node variable, and the equation contains all the preceding causal node variables of that node, as well as the node's own error term, used to describe the generation mechanism and causal dependencies of the node variables.

[0063] Preferably, the system can use numerical methods to solve the structural equations, and optimize the equation parameters through iterative calculations to obtain a causal structure function that accurately characterizes the generation mechanism and causal dependency of each node variable. The causal structure function clarifies the quantitative relationship between each node variable and other causal node variables. Based on the obtained causal structure function, the system uses the average causal effect analysis method to calculate the causal effect between any two nodes after intervention. That is, keeping other node variables constant, changing the value of one node variable, and calculating the change in the other node variable, thus obtaining the average causal effect between the two nodes. The system arranges the average causal effect values ​​between all nodes in a matrix according to the node number order to generate a causal effect matrix. The rows and columns of the causal effect matrix correspond to the monitoring nodes of the distribution network, and the elements in the matrix correspond to the average causal effect intensity between two nodes. Through this causal effect matrix, the system achieves a quantitative characterization of the causal effect between nodes in the distribution network.

[0064] S4: Construct a causal augmented graph neural network based on dynamic causal graphs and causal effect matrices. Use dynamic causal graphs as the network graph structure and embed the causal effect matrix as the causal weights for message passing into the network propagation process. Use historical data from multi-source tensors with fault type and location as supervision labels for training to generate a diagnostic model that integrates causal prior knowledge.

[0065] Specifically, the system constructs a causal-enhanced graph neural network based on a dynamic causal graph and a causal effect matrix. The graph neural network possesses powerful topological feature extraction capabilities, effectively handling the correlations between nodes in the distribution network. The system uses the dynamic causal graph as the basic network graph structure of the graph neural network, ensuring that the nodes and edges of the graph neural network directly correspond to the nodes and causal relationship edges in the distribution network, guaranteeing that the network structure is consistent with the actual causal topology of the distribution network. The system uses the causal effect matrix as causal weights in the message passing process, embedding it into the feature propagation and aggregation stages of the graph neural network. During feature transmission, the system adjusts the feature transmission weights between different nodes based on the values ​​in the causal effect matrix, giving nodes with high causal effect strength stronger feature transmission capabilities. This allows the feature learning process of the graph neural network to integrate the causal prior knowledge of the distribution network, breaking away from the purely data-driven statistical correlation dependence of traditional graph neural networks and improving the accuracy and interpretability of the network's fault feature identification. The system uses historical time-series data from multi-source tensors as training input and selects fault types and fault locations from historical fault data of the distribution network as supervision labels for iterative training of the causal-enhanced graph neural network. During training, the system inputs historical data into the network, obtains fault diagnosis prediction results through forward propagation, compares them with supervision labels, calculates the prediction error, and uses the backpropagation algorithm to update the network's weights and bias parameters, continuously optimizing network performance. After multiple rounds of iterative training, when the network's prediction error reaches a stable state, the system stops training and generates a distribution network fault diagnosis model that integrates physical causal prior knowledge.

[0066] S5: Based on the causal structure function, counterfactual intervention is performed on the confounding factors in the samples to be diagnosed obtained from the real-time data stream to generate a counterfactual sample set. The samples to be diagnosed and the counterfactual sample set are respectively input into the diagnostic model for probabilistic diagnosis. The fusion weight is determined by combining the effect intensity of the confounding factors in the causal effect matrix to perform weighted fusion of the diagnostic probabilities and generate the final diagnostic result after removing the interference of confounding factors.

[0067] Specifically, the system acquires the samples to be diagnosed from the real-time data stream. These samples include quasi-steady-state measurement data, dynamic synchronization phasor data, and meteorological data from the real-time operation of the distribution network, consistent with the data types acquired in step S1. Based on the causal structure function obtained in step S3, the system performs counterfactual intervention processing on the samples to be diagnosed. The core of the counterfactual intervention processing is to remove confounding factors from the samples to be diagnosed. The system identifies confounding factors in the samples to be diagnosed, including meteorological interference, load fluctuations, and measurement noise. These factors can cause false correlations between fault characteristics and non-fault characteristics, affecting the accuracy of diagnosis. The system sets the confounding factors to an interference-free baseline state, i.e., eliminating the influence of confounding factors on the distribution network operation data. Based on the causal structure function, it generates counterfactual samples under this baseline state. By repeatedly adjusting the interference state of the confounding factors, a set of counterfactual samples covering multiple confounding scenarios is generated.

[0068] Furthermore, the system inputs the original samples to be diagnosed and the counterfactual sample set into the diagnostic model trained in step S4. The diagnostic model performs fault diagnosis on each sample and outputs the diagnostic probability distribution corresponding to the fault type and fault location. The system extracts the causal effect strength between confounding factors and fault features in the causal effect matrix, and determines the fusion weight of each diagnostic result based on this strength. The higher the causal effect strength, the greater the weight of the corresponding diagnostic result, thereby highlighting the diagnostic results corresponding to true causal associations and suppressing false diagnostic results caused by confounding factors. The system adopts a weighted fusion method to integrate and calculate the diagnostic probabilities of the original samples to be diagnosed and the counterfactual sample set, eliminating the interference of false associations caused by confounding factors, and finally generating accurate fault diagnosis results, clarifying the fault type and fault location of the distribution network, and providing a reliable basis for distribution network fault handling.

[0069] In summary, the intelligent fault diagnosis method for distribution networks provided in this application can accurately capture the real information transmission paths between nodes during the operation of the distribution network by constructing a dynamic causal graph using the transfer entropy algorithm, effectively filtering out spurious correlations caused by meteorological interference and load fluctuations. The introduction of a causal effect matrix quantified by structural equation modeling transforms abstract causal strength into quantifiable computational weights, providing a physically interpretable guiding basis for subsequent message transmission in the neural network. Embedding the causal effect matrix as prior knowledge into the message transmission process of the graph neural network guides the model to focus on the features of neighboring nodes that truly have causal influence, avoiding the attention mechanism being misled by noisy data. Based on the counterfactual sample set generated by the causal structure function, the system state under different scenarios can be simulated by intervening in confounding factors, enabling robust verification of diagnostic results under environmental changes. Combining the causal effect strength to determine the fusion weights and weightedly fusing the probabilities of the original diagnosis and counterfactual diagnosis can effectively suppress the risk of misjudgment caused by changes in operating mode or sudden changes in the external environment, thereby achieving the model's adaptive adjustment capability under multiple operating conditions and ultimately improving the credibility and interpretability of the diagnostic results.

[0070] In one embodiment, S1 of the intelligent fault diagnosis method for power distribution networks provided by the present invention specifically includes the following steps:

[0071] S11: Acquire quasi-steady-state measurement data of each node in the distribution network, dynamic synchronous phasor data of each PMU deployment location, and meteorological data of each meteorological monitoring station. Perform missing value imputation and outlier removal processing on various types of data to generate a cleaned multi-source dataset.

[0072] Specifically, the system acquires quasi-steady-state measurement data from each node of the distribution network, dynamic synchronization phasor data from each PMU deployment location, and meteorological data from each meteorological monitoring station. The quasi-steady-state measurement data from each node is collected through SCADA terminals deployed at each monitoring node of the distribution network, covering all monitoring nodes and encompassing various electrical parameters during node operation. The dynamic synchronization phasor data from each PMU deployment location is collected through distributed PMU measurement devices. Each PMU deployment corresponds to a specific monitoring node in the distribution network, and the collected data covers high-frequency dynamic electrical parameters such as synchronization phasors, frequency, and phase angle of that node, with a collection frequency higher than that of the quasi-steady-state measurement data. The meteorological data from each meteorological monitoring station is collected through distributed meteorological monitoring terminals within the distribution network coverage area. Each meteorological monitoring station corresponds to a specific area of ​​the distribution network, and the collected data covers various meteorological parameters for that area, reflecting the environmental conditions of the distribution network operation.

[0073] Due to factors such as the operational status of the data acquisition terminal, communication transmission processes, and external environmental interference, missing and outlier values ​​may appear in various types of acquired data. Missing values ​​originate from data acquisition terminal malfunctions, communication interruptions, etc., while outliers arise from acquisition noise, equipment false alarms, etc. Such data can affect the accuracy of subsequent processing. Therefore, the system performs missing value imputation and outlier removal on various types of data. The system adopts a missing value imputation method adapted to time-series data. Based on the variation pattern of the time-series sequence containing the missing data, imputation is performed through the correlation relationship between adjacent valid data to ensure that the imputed data conforms to the actual operation pattern of the distribution network. The system analyzes the distribution characteristics of various types of data to identify data exceeding the normal operating range as outliers. Statistical methods are used to remove the identified outliers, while retaining valid data in the dataset to avoid the accidental deletion of valid data. After missing value imputation and outlier removal, the system integrates various types of data to generate a cleaned multi-source dataset. The cleaned multi-source dataset includes quasi-steady-state measurement data, dynamic synchronous phasor data, and meteorological data. The dimensions of the various types of data are different. The voltage, current, power, and other parameters in the quasi-steady-state measurement data have different dimensions. The dimensions of the phasor, frequency, phase angle, and other parameters in the dynamic synchronous phasor data are also inconsistent. The various parameters in the meteorological data also have different dimensions.

[0074] S12: Normalize the various types of data in the cleaned multi-source dataset, map features of different dimensions to a unified numerical range, and generate a standardized feature sequence.

[0075] Specifically, features with different dimensions can lead to significant differences in data numerical ranges. Directly using these features for subsequent processing can result in unbalanced feature weights, affecting the accuracy of the processing results. Therefore, the system needs to normalize various types of data. The system employs a standardized normalization method, processing each type of feature data in the multi-source dataset separately. By calculating the overall distribution characteristics of this type of feature data, the system determines the normalization transformation rules. Following these rules, the system maps various types of feature data to a unified numerical range, eliminating the influence of different dimensions and ensuring that all types of feature data have the same numerical magnitude. During the normalization process, the system preserves the original variation patterns of various types of data and does not alter the relative correlations between data, ensuring that the normalized feature data accurately reflects the operating status and environmental conditions of the power distribution network. After normalization, the system converts various types of data into standardized feature sequences. Each feature sequence corresponds to the time-series variation data of a type of parameter. The numerical ranges of different feature sequences are consistent. The standardized feature sequences contain different types of time-series data. Quasi-steady-state measurement data, dynamic synchronous phasor data, and meteorological data are collected at different frequencies. The dynamic synchronous phasor data collected by PMU deployment has the highest collection frequency, followed by the quasi-steady-state measurement data collected by SCADA terminals, and the meteorological data collected by meteorological monitoring terminals has the lowest collection frequency.

[0076] S13: Perform time axis alignment processing on the feature sequence, calculate the optimal alignment path between data sequences with different sampling frequencies, eliminate the phase difference caused by communication delay through path mapping, and generate time-aligned multi-source data.

[0077] Specifically, different sampling frequencies lead to inconsistent timestamp distributions across various feature sequences. Simultaneously, communication delays cause phase differences in different types of data collected at the same time. Timestamp misalignment and phase differences prevent precise matching of multi-source data, affecting the accuracy of subsequent tensor organization. Therefore, the system needs to perform timeline alignment on the feature sequences. The system employs a timeline alignment algorithm to analyze feature sequences with different sampling frequencies and calculate the optimal alignment path between various data sequences. By comparing the timestamp distributions of different feature sequences and considering the temporal variation patterns of the data, the system determines the alignment path that maximizes the similarity of various data sequences. This path accurately matches the time nodes of data from different sampling frequencies. The system performs path mapping using the optimal alignment path, mapping feature sequences with different sampling frequencies onto a unified timeline. Simultaneously, based on the path mapping relationship, it eliminates phase differences caused by communication delays, ensuring precise matching of various feature data corresponding to the same time node and guaranteeing consistency of different data types in the time dimension. After timeline alignment, the system generates time-aligned multi-source data, achieving synchronization of various data types in the time dimension, eliminating the impact of time misalignment and phase differences, and providing a time-synchronized data foundation for subsequent tensor organization.

[0078] S14: Organize the multi-source data into a three-dimensional multi-source tensor by tensing according to three dimensions: timestamp, node number, and feature type.

[0079] Specifically, the system tensors the time-aligned multi-source data according to three dimensions: timestamp, node number, and feature type, constructing a three-dimensional multi-source tensor. The time-aligned multi-source data has achieved synchronization in the time dimension, including quasi-steady-state measurement data from each node of the distribution network, dynamic synchronization phasor data from each PMU deployment point, and meteorological data from each meteorological monitoring station. All types of data have clear timestamps, corresponding node numbers, and feature type identifiers. The system uses timestamps as the first dimension, sorting all data according to the order of timestamps. Each timestamp corresponds to a time node, covering all data acquisition times and ensuring that the time dimension fully reflects the temporal changes of the data. The system uses node numbers as the second dimension, corresponding to all monitoring nodes in the distribution network, including SCADA terminal monitoring nodes, PMU deployment nodes, and distribution network area nodes corresponding to meteorological monitoring stations. Each node number corresponds to a set of all feature data for that node, ensuring that the spatial dimension fully covers the entire monitoring range of the distribution network.

[0080] Furthermore, the system uses feature type as a third dimension. Feature type encompasses various specific parameters from quasi-steady-state measurement data, dynamic synchronous phasor data, and meteorological data. Each feature type corresponds to the time-series data of a specific parameter, ensuring that the feature dimension fully includes the electrical and environmental characteristics of the power distribution network operation. Based on this three-dimensional division, the system integrates and arranges the time-aligned multi-source data, using each timestamp, node number, and specific data corresponding to each feature type as elements of a tensor to construct a three-dimensional multi-source tensor. This three-dimensional multi-source tensor achieves integrated organization of multi-source data across time, space, and features, transforming dispersed multi-source data into a structured and standardized tensor form.

[0081] In one embodiment, S2 of the intelligent fault diagnosis method for distribution networks provided by the present invention specifically includes the following steps:

[0082] S21: Perform sliding window partitioning on the voltage and current feature sequences of each node in the multi-source tensor, calculate the mutual information value between any two node feature sequences within each window, and construct an initial adjacency matrix based on the mutual information value.

[0083] Specifically, based on the constructed three-dimensional multi-source tensor, the system performs sliding window partitioning on the voltage and current feature sequences of each node. These voltage and current feature sequences are core electrical characteristics reflecting the operating state of distribution network nodes. The operating state of the distribution network exhibits dynamic evolution, with varying correlation strengths between nodes across different time periods. Sliding window partitioning captures this dynamic change. The system partitions the voltage and current feature sequences of each node according to a fixed window length and sliding step size, dividing the continuous time-series feature sequences into multiple overlapping window segments, each corresponding to a continuous operating time period. After partitioning, the system calculates the mutual information value between any two node feature sequences within each window. This mutual information value quantifies the correlation strength between two random variables, adapting to the nonlinear characteristics of distribution network operation. Preferably, the mutual information value can be calculated using the following formula:

[0084]

[0085] Where X and Y represent the feature sequences of two different nodes, and These represent the marginal probability distributions of X and Y, respectively. Represents the joint probability distribution of X and Y. The mutual information value represents the feature sequences of two nodes. The system uses each node of the distribution network as the row and column of a matrix, and uses the mutual information value of any two nodes within each window as the element of the corresponding position in the matrix to construct the initial adjacency matrix for each window.

[0086] S22: Based on the initial adjacency matrix, perform directional calculation on the information transmission direction between each pair of nodes, determine the dominant direction of information flow by comparing the transmission entropy values ​​in different directions, and generate a directed graph structure with directional weights.

[0087] Specifically, based on the constructed initial adjacency matrix, the system performs directional calculations on the information transmission direction between each pair of nodes. The initial adjacency matrix only reflects the strength of the association between nodes and cannot distinguish the direction of the association. Causal relationships have a clear directionality; therefore, the direction of information transmission between nodes must be determined to construct a directed association structure that conforms to the actual causal mechanism. The system calls the transfer entropy algorithm to calculate the bidirectional information transmission strength for each pair of nodes corresponding to each non-zero element in the initial adjacency matrix. The transfer entropy value is calculated using the following formula:

[0088]

[0089] Where TE(X→Y) represents the transfer entropy value from node X to node Y. and These represent the feature values ​​of node X at time t and time t+1, respectively. Let represent the feature value of node Y at time t, and p represent the joint probability distribution of the corresponding variables. The system compares the bidirectional transmission entropy values ​​of each node pair to determine the dominant direction of information flow. The dominant transmission direction is taken as the direction of the directed edge, and the transmission entropy value in that direction is taken as the weight of the directed edge. The directed edge information of all node pairs is integrated to generate a directed graph structure with directional weights.

[0090] S23: Perform significance testing on each directed edge in the directed graph structure, filter out weakly related edges with a propagation entropy value lower than the preset confidence level, retain causal relationship edges with statistical significance, and generate a dynamic causal graph.

[0091] Specifically, the system performs significance testing on each directed edge in the directed graph structure, removing weakly correlated edges to retain those with statistically significant causal relationships, thereby generating a dynamic causal graph. Some directed edges in the directed graph structure have low transit entropy values, originating from irrelevant factors such as random noise and normal load fluctuations, lacking true causal significance and interfering with the accuracy of subsequent causal analysis; therefore, significance testing is necessary. The system employs hypothesis testing, setting a pre-defined confidence level as the standard for judging whether a directed edge is statistically significant. The significance test uses the formula:

[0092]

[0093] Where S represents the test statistic, and TE represents the transitive entropy value corresponding to the directed edge. This represents the mean propagation entropy generated by random noise. represents the standard deviation of the propagation entropy of random noise, and n represents the number of samples. The system calculates the statistic corresponding to the propagation entropy value of each directed edge, compares the statistic with the critical value corresponding to the preset confidence level, filters out weakly related edges with statistics below the critical value, and retains causal relationship edges with statistics above the critical value and statistical significance. After integration, a dynamic causal graph is generated, providing reliable causal topology support for subsequent causal effect quantification and diagnostic model construction.

[0094] In one embodiment, such as Figure 2 As shown, step S3 of the intelligent fault diagnosis method for power distribution networks provided by this invention specifically includes the following steps:

[0095] S31: Based on the set of parent nodes of each node in the dynamic causal graph, structural equation modeling is performed on the mapping relationship between the state features of each node and the features of the parent nodes. Basis function expansion and fitting are performed on the mapping relationship between the state features and the features of the parent nodes. The basis function weight coefficients are estimated by the least squares method to generate the initial structure function of each node.

[0096] Specifically, the system performs preliminary construction of the causal structure function for each node based on the set of parent nodes in the dynamic causal graph. First, the system extracts the set of parent nodes corresponding to each node in the dynamic causal graph; the set of parent nodes consists of all preceding nodes that have a direct causal relationship with that node. For each node, the system models the mapping relationship between its state features and the features of each parent node in the set of parent nodes using structural equation modeling. The structural equation is used to characterize the quantitative dependency relationship between the node's state features and the features of its parent nodes. The system then performs basis function expansion fitting on the mapping relationship between the state features and the parent node features. Basis function expansion can transform complex nonlinear mapping relationships into a solvable linear combination form. The fitting uses the following formula:

[0097]

[0098] in, This is the initial structure function prototype for node i. These are the basis function weight coefficients. For the selected basis functions, Let be the set of parent nodes of node i. The system estimates the basis function weight coefficients using the least squares method. By minimizing the error between the actual state characteristics of the nodes and the fitted values, the optimal weight coefficients are obtained. Substituting the optimal weight coefficients into the basis function expansion generates the initial structure function for each node.

[0099] S32: Perform additivity decomposition on the initial structure function to separate the contributions of exogenous and endogenous variables. Iteratively optimize the structure function parameters by minimizing the prediction error to generate the causal structure function for each node.

[0100] Specifically, the system performs additivity decomposition on the initial structure function of each node. Additivity decomposition is used to separate the contributions of exogenous disturbance variables and endogenous variables to the node state features. Exogenous disturbance variables are those that affect the node state but are not affected by other nodes, while endogenous variables are those that are affected by the parent node and participate in the composition of the node state features. Through additivity decomposition, the system splits the initial structure function into two parts: the contribution of endogenous variables and the contribution of exogenous disturbance variables, clarifying the independent influence of the two types of variables on the node state features. With the goal of minimizing prediction error, the system iteratively optimizes the structure function parameters. During the iteration process, the parameters of the contribution of endogenous variables and the contribution of exogenous disturbance variables are continuously adjusted, and the prediction error after each parameter adjustment is calculated until the prediction error reaches a stable state. After multiple rounds of iterative optimization, the system determines the optimal parameter combination, substitutes the optimal parameters into the decomposed structure function, and generates the causal structure function for each node. This function can accurately characterize the quantitative causal relationship between the node state features, the parent node features, and the exogenous disturbance variables.

[0101] S33: Based on the causal structure function, a forward propagation calculation is performed after applying a unit impulse perturbation to each node. After applying a unit impulse perturbation to the causal structure function of each node, a forward inference process is performed. By comparing the influence of the changes in the characteristics of the parent node set before and after the impulse is applied on the characteristics of the child node, a causal effect matrix is ​​generated by traversing all node pairs.

[0102] Specifically, the system generates a causal effect matrix based on the causal structure function of each node. The system applies a unit impulse perturbation to each node, which simulates minute changes in the node's state. Then, it performs forward propagation calculations and forward deduction on the perturbation-enhanced causal structure function to track the transmission of the perturbation signal between nodes. The system compares the impact of changes in the parent node set characteristics on the child node characteristics before and after the unit impulse application; this magnitude is used to quantify the causal effect strength of the parent node on the child node. The system iterates through all node pairs, repeating the above perturbation, deduction, and comparison process, and calculates the modified causal effect strength of each node pair using a preset formula. The formula for calculating the causal effect matrix is:

[0103]

[0104] in, Let be the strength of the modified causal effect of node j on node i. Let i be the causal structure function of node i. Let be the set of parent nodes of node i at time t. Let represent the state characteristics of node j at time t. For time-varying decay factor control parameters, Let j be the information propagation delay from node j to node i. Let the oscillation characteristic period between node j and node i be . Let t be the phase difference between node j and node i at time t, and T be the total time length.

[0105] In one embodiment, S4 of the intelligent fault diagnosis method for power distribution networks provided by the present invention specifically includes the following steps:

[0106] S41: Based on the topological structure of the dynamic causal graph, construct the adjacency matrix of the graph neural network, embed the elements in the causal effect matrix as the initial causal weights for message passing into the corresponding positions of the adjacency matrix, perform causal effect normalization on the weight values ​​of each edge in the adjacency matrix, and construct a graph neural network framework that integrates causal weights.

[0107] Specifically, the system constructs an adjacency matrix for a graph neural network based on the topology of a dynamic causal graph. The nodes of the dynamic causal graph correspond to monitoring nodes in the distribution network, and directed edges correspond to causal relationships between nodes. The system directly maps this topology to the basic adjacency matrix of the graph neural network. The rows and columns of the adjacency matrix correspond to nodes in the distribution network, and matrix elements indicate whether causal relationships exist between nodes. The system extracts elements from the causal effect matrix and uses them as initial causal weights for message passing, embedding them in the corresponding positions of the adjacency matrix. This ensures that the adjacency matrix not only contains relationships between nodes but also incorporates quantified causal effect strength. The system performs causal effect normalization on the weight values ​​of each edge in the adjacency matrix to eliminate the influence of causal effect weights of different magnitudes. The normalization uses the following formula:

[0108]

[0109] in, The normalized edge weights Let N be the modified causal effect strength of node j on node i, and N be the total number of nodes in the distribution network. After normalization, the system integrates the adjacency matrix and the normalized causal weights to construct a graph neural network framework that integrates causal weights.

[0110] S42: Perform multi-head attention calculation on the hidden states of each node in the graph neural network framework, and concatenate and fuse the node's own features, neighboring node features and corresponding causal effect weights. Adaptively adjust the contribution of different causal paths through learnable attention parameters to generate causally enhanced node representations.

[0111] Specifically, the system performs multi-head attention computation on the hidden states of each node in the graph neural network framework. Multi-head attention is used to adaptively capture the importance of different causal paths. The system first extracts the self-features and neighbor node features of each node. The neighbor node features are determined based on the topological relationships of the dynamic causal graph, i.e., the features corresponding to the node's parent and child nodes. The system concatenates and fuses the node's self-features, neighbor node features, and corresponding normalized causal effect weights to form a concatenated feature vector. The system then performs multi-head attention computation on the concatenated feature vector using learnable attention parameters, calculated using the following formula:

[0112]

[0113] Where Q is the query vector, corresponding to the node's own features; K is the key vector, corresponding to the features of neighboring nodes; and V is the value vector, corresponding to the causal effect weights. The key vector dimension is used. Through this calculation, the system adaptively adjusts the contribution of different causal paths, strengthens the influence of strong causal paths, weakens the interference of weak causal paths, and finally generates a causally enhanced node representation that integrates the node's own features, neighbor features, and causal prior knowledge.

[0114] S43: The graph neural network framework is trained end-to-end based on historical data from multi-source tensors. Fault type labels and fault location labels are used as supervision signals for backpropagation optimization. The network parameters are iteratively updated through gradient descent algorithm until the loss function converges, generating a diagnostic model that integrates causal prior knowledge.

[0115] Specifically, the system performs end-to-end training on the constructed graph neural network framework based on historical data from multi-source tensors. The system extracts historical time-series data from the multi-source tensors as input data for the graph neural network. This data covers multi-source features under normal operation and various fault states of the distribution network. The system selects fault type labels and fault location labels from the historical fault data of the distribution network as supervision signals for the training process, used to measure the model's diagnostic error. The system employs the backpropagation algorithm, combined with the supervision signals, for parameter optimization, iteratively updating all learnable parameters of the graph neural network, including attention parameters and weight coefficients, through the gradient descent algorithm. During training, the system calculates the loss value between the model's output diagnostic results and the supervision labels. The loss value is calculated using the formula:

[0116]

[0117] in, Here, M represents the loss function value, and M represents the total number of samples. For the true labels of the samples, The system predicts labels for the model. It continuously iterates and updates parameters until the loss function reaches a stable convergence state, at which point training stops, generating a diagnostic model that integrates causal prior knowledge. This model can leverage causal correlation knowledge to improve the accuracy and reliability of fault diagnosis.

[0118] In one embodiment, S5 of the intelligent fault diagnosis method for power distribution networks provided by the present invention specifically includes the following steps:

[0119] S51: Perform confounding factor identification processing on the samples to be diagnosed obtained from the real-time data stream, mark meteorological features and load fluctuation features as confounding variables to be intervened, and extract the state features of each node in the samples to be diagnosed as feature benchmark values ​​for counterfactual reasoning.

[0120] Specifically, the system performs confounding factor identification processing on the samples to be diagnosed acquired from the real-time data stream. The samples contain various characteristic data of the real-time operation of the distribution network. By analyzing the correlation between characteristic data and fault characteristics, the system identifies confounding factors that interfere with fault diagnosis. The system marks meteorological characteristics and load fluctuation characteristics as confounding variables to be intervened in, as these characteristics have spurious correlations with fault characteristics and are prone to misjudgment by the diagnostic model. The system extracts the state characteristics of each node in the samples to be diagnosed, including core electrical characteristics such as voltage and current, and uses these as feature benchmark values ​​for counterfactual reasoning. These feature benchmark values ​​are used as the basis for comparison in subsequent intervention simulations. The system uses a confounding factor identification formula to assist in identification; the calculation formula is as follows:

[0121]

[0122] in, Let be the confounding interference coefficient of the j-th type feature. Let be the value of the j-th feature at time t. Let t be the fault identification variable at time t, cov be the covariance calculation function, T be the sample time length, and features with interference coefficients higher than the set standard are marked as confounding variables.

[0123] S52: Based on the causal structure function, different combinations of intervention values ​​are applied sequentially to the confounding variables in the feature baseline values, while the non-confounding variables are kept unchanged. The forward inference of the state features of each node after the intervention is calculated through the causal structure function, generating multiple sets of counterfactual state features after the intervention.

[0124] Specifically, based on a causal structure function, the system sequentially applies different combinations of intervention values ​​to confounding variables in the feature baseline values. The system keeps unconfounding variables constant; these unconfounding variables are feature variables with a genuine causal relationship to the fault and no interfering effect, ensuring that the intervention process only changes the values ​​of confounding variables and does not affect other core features. The system sequentially substitutes different combinations of intervention values ​​into the causal structure function of each node, performs forward inference calculations through the causal structure function, tracks the response changes of each node's state features after changes in the values ​​of confounding variables, and obtains the values ​​of each node's state features under different intervention scenarios. Preferably, the forward inference uses the following formula:

[0125]

[0126] in, Let i be the state characteristic value of node i at time t after intervention. Let i be the causal structure function of node i. The features of the non-hybrid parent node are fixed and unchanging. This represents the intervention value for confounding variables. The system integrates node state features from all intervention scenarios to generate multiple sets of counterfactual state features after intervention.

[0127] S53: Reorganize the counterfactual state features according to the dimensional organization of multi-source tensors to construct a counterfactual sample set containing multiple counterfactual samples.

[0128] Specifically, the system reorganizes the generated counterfactual state features according to the dimensional organization of multi-source tensors to construct a counterfactual sample set. The multi-source tensor dimensional organization consists of three dimensions: timestamp, node number, and feature type. The system follows this organization rule to perform dimensional matching for each set of counterfactual state features. The system assigns a corresponding timestamp to each set of counterfactual state features, ensuring consistency with the timestamp of the sample to be diagnosed, thus guaranteeing the uniformity of the time dimension. The system maps the counterfactual state features to each node according to the distribution network node number, ensuring the accuracy of the node dimension. The system categorizes the counterfactual state features by feature type, covering various features such as voltage, current, weather, and load fluctuations, ensuring the completeness of the feature dimensions. The system treats each reorganized set of counterfactual state features as a counterfactual sample, integrates all counterfactual samples, and constructs a counterfactual sample set containing multiple counterfactual samples, providing standardized input for subsequent parallel diagnostic calculations.

[0129] S54: Input the sample to be diagnosed into the diagnostic model for forward propagation calculation, and map it layer by layer through graph convolutional layers and fully connected layers to output the original diagnostic probability vector of each node belonging to various types of faults.

[0130] Specifically, the system inputs the sample to be diagnosed into the trained diagnostic model for forward propagation computation. After preprocessing, the sample is converted according to the input format of the graph neural network and input into the graph convolutional layer of the model. The graph convolutional layer extracts and aggregates features from the node features and causal weights in the sample, achieving spatial propagation of node features through graph convolution operations. The graph convolution calculation uses the following formula:

[0131]

[0132] in, For the feature output of node i in the (l+1)th layer, Let i be the set of neighboring nodes. To normalize causal weights, For the features of node j in the l-th layer, The learnable weights of the l-th layer, is the activation function. After the features are processed by the graph convolutional layer, they are input to the fully connected layer and mapped layer by layer, finally outputting the original diagnostic probability vector of each node belonging to various types of faults. Each element in the probability vector corresponds to the diagnostic probability of a fault type.

[0133] S55: Input each sample in the counterfactual sample set into the diagnostic model in sequence for parallel computation, and obtain the corresponding counterfactual diagnostic probability vector set.

[0134] Specifically, the system sequentially inputs each sample from the counterfactual sample set into the diagnostic model for parallel computation. The system employs a parallel scheduling mechanism, allocating samples from the counterfactual sample set to different computational units, performing diagnostic calculations simultaneously to improve computational efficiency. The computation process for each counterfactual sample is consistent with that of the sample to be diagnosed, undergoing layer-by-layer processing through graph convolutional layers and fully connected layers to output the corresponding counterfactual diagnostic probability vector. The counterfactual diagnostic probability vector reflects the diagnostic probability of each node belonging to a specific type of fault after removing corresponding profanity. The system collects the diagnostic probability vectors of all counterfactual samples and organizes them into a counterfactual diagnostic probability vector set, which contains multiple sets of diagnostic probabilities under different profanity intervention scenarios.

[0135] S56: Based on the mean effect intensity of confounding variables on the voltage and current characteristics of each node in the causal effect matrix, calculate the fusion weight coefficient, and perform weighted averaging on the original diagnostic probability vector and the counterfactual diagnostic probability vector set according to the fusion weight coefficient to generate the final diagnostic result after removing the interference of confounding factors.

[0136] Specifically, the system calculates the fusion weighting coefficients based on the mean effect strength of confounding variables on the voltage and current characteristics of each node in the causal effect matrix. The system extracts the effect strength between the node corresponding to the confounding variable and each fault node in the causal effect matrix and calculates its mean over the time dimension. This mean is used to measure the degree of interference of the confounding variables on the fault diagnosis of each node. Preferably, the fusion weighting coefficients are calculated using the following formula:

[0137]

[0138] in, Let be the fusion weight coefficient of the k-th counterfactual sample. Let K be the mean effect strength of the k-th confounding variable on node i, and K be the number of counterfactual samples. The system performs a weighted average of the original diagnostic probability vector and the counterfactual diagnostic probability vector set according to the fusion weight coefficient. The weighted average result is the final diagnostic result after removing the interference of confounding factors, which clarifies the fault type and corresponding probability of each node.

[0139] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0140] Based on the same inventive concept, this application also provides a distribution network fault intelligent diagnosis system for implementing the above-mentioned distribution network fault intelligent diagnosis method. The solution provided by this system is similar to the implementation scheme described in the above method; therefore, the specific limitations in one or more distribution network fault intelligent diagnosis system embodiments provided below can be found in the limitations of the distribution network fault intelligent diagnosis method described above, and will not be repeated here.

[0141] Preferably, such as Figure 3 As shown, the present invention provides an intelligent fault diagnosis system 600 for power distribution networks, which is configured with the following modules:

[0142] The multi-source data spatiotemporal alignment module 610 is used to perform spatiotemporal alignment processing on the acquired quasi-steady-state measurement data of the distribution network, dynamic synchronous phasor data and meteorological data using a dynamic time warping algorithm, and to construct a multi-source tensor containing three dimensions: time, node and feature.

[0143] The dynamic causal graph construction module 620 is used to calculate the information transmission direction and intensity between feature sequences of any two nodes based on multi-source tensors and the transfer entropy algorithm, and to perform significance filtering on the calculated information transmission intensity to select causal relationship edges with statistical significance, thereby generating a dynamic causal graph.

[0144] The causal effect quantification module 630 is used to quantify the causal relationship of each node based on the dynamic causal graph and multi-source tensor, and introduce the structural equation model. It obtains the causal structure function of each node by solving the structural equation, and calculates the average causal effect between nodes based on the causal structure function to generate the causal effect matrix.

[0145] The Causal Augmented Graph Network Training Module 640 is used to construct a causal augmented graph neural network based on a dynamic causal graph and a causal effect matrix. The dynamic causal graph is used as the network graph structure and the causal effect matrix is ​​used as the causal weights for message passing and embedded in the network propagation process. The model is trained using historical data from multi-source tensors with fault type and location as supervision labels to generate a diagnostic model that integrates causal prior knowledge.

[0146] The counterfactual diagnostic fusion module 650 is used to perform counterfactual intervention processing on the confounding factors in the samples to be diagnosed obtained from the real-time data stream based on the causal structure function, generate a counterfactual sample set, input the samples to be diagnosed and the counterfactual sample set into the diagnostic model for probability diagnosis, and determine the fusion weight by combining the effect intensity of the confounding factors in the causal effect matrix to perform weighted fusion of the diagnostic probabilities, and generate the final diagnostic result after removing the interference of confounding factors.

[0147] Preferably, the multi-source data spatiotemporal alignment module 610 provided in this application is configured with the following units:

[0148] The multi-source data cleaning unit is used to acquire quasi-steady-state measurement data of each node of the distribution network, dynamic synchronous phasor data of each PMU deployment location, and meteorological data of each meteorological monitoring station. It performs missing value imputation and outlier removal on various types of data to generate a cleaned multi-source dataset.

[0149] The feature normalization processing unit is used to normalize various types of data in the cleaned multi-source dataset, mapping features of different dimensions to a unified numerical range and generating a standardized feature sequence.

[0150] The dynamic time axis alignment unit is used to perform time axis alignment processing on feature sequences, calculate the optimal alignment path between data sequences with different sampling frequencies, eliminate the phase difference caused by communication delay through path mapping, and generate time-aligned multi-source data.

[0151] The three-dimensional multi-source tensor construction unit is used to tensile and organize multi-source data according to three dimensions: timestamp, node number, and feature type, to construct a three-dimensional multi-source tensor.

[0152] Preferably, the dynamic cause-effect graph construction module 620 provided in this application is configured with the following units:

[0153] The initial adjacency matrix construction unit is used to perform sliding window partitioning on the voltage and current feature sequences of each node in the multi-source tensor, calculate the mutual information value between any two node feature sequences within each window, and construct the initial adjacency matrix based on the mutual information value.

[0154] The transfer entropy orientation calculation unit is used to perform orientation calculation on the information transfer direction between each pair of nodes based on the initial adjacency matrix. By comparing the transfer entropy values ​​in different directions, the dominant direction of information flow is determined, and a directed graph structure with directional weights is generated.

[0155] The saliency screening graph construction unit is used to perform saliency testing on each directed edge in the directed graph structure, filtering out weakly associated edges with a transit entropy value lower than a preset confidence level, retaining causal relationship edges with statistical significance, and generating a dynamic causal graph.

[0156] Preferably, the causal effect quantification module 630 provided in this application is configured with the following units:

[0157] The initial structural equation modeling unit is used to model the mapping relationship between the state features of each node and the features of its parent nodes based on the set of parent nodes of each node in the dynamic causal graph. It performs basis function expansion and fitting on the mapping relationship between the state features and the features of its parent nodes, estimates the basis function weight coefficients by least squares method, and generates the initial structural function of each node.

[0158] The causal structure function optimization unit is used to perform additivity decomposition on the initial structure function, separate the contributions of exogenous perturbation variables from those of endogenous variables, and iteratively optimize the structure function parameters by minimizing the prediction error to generate the causal structure function for each node.

[0159] The impulse causality effect calculation unit is used to perform forward propagation calculations after applying a unit impulse perturbation to each node based on the causal structure function, and to perform forward inference processing after applying a unit impulse perturbation to the causal structure function of each node. By comparing the influence of the changes in the characteristics of the parent node set before and after the impulse is applied on the characteristics of the child node, the unit traverses all node pairs to generate the causal effect matrix.

[0160] Preferably, the causal augmentation graph network training module 640 provided in this application is configured with the following units:

[0161] The causal weight graph network building unit is used to construct the adjacency matrix of the graph neural network based on the topology of the dynamic causal graph. The elements in the causal effect matrix are embedded into the corresponding positions of the adjacency matrix as the initial causal weights for message passing. The weight values ​​of each edge in the adjacency matrix are normalized by causal effect to construct a graph neural network framework that integrates causal weights.

[0162] The causal enhancement node representation unit is used to perform multi-head attention calculation on the hidden state of each node in the graph neural network framework. It concatenates and fuses the node's own features, the features of its neighboring nodes, and the corresponding causal effect weights. It adaptively adjusts the contribution of different causal paths through learnable attention parameters to generate causal enhancement node representations.

[0163] The diagnostic model end-to-end training unit is used to perform end-to-end training of the graph neural network framework based on historical data from multi-source tensors. It uses fault type labels and fault location labels as supervision signals for backpropagation optimization, and iteratively updates the network parameters through the gradient descent algorithm until the loss function converges, generating a diagnostic model that integrates causal prior knowledge.

[0164] Preferably, the counterfactual diagnostic fusion module 650 provided in this application is configured with the following units:

[0165] The confounding factor identification and labeling unit is used to identify confounding factors in the samples to be diagnosed obtained from the real-time data stream, mark meteorological features and load fluctuation features as confounding variables to be intervened, and extract the state features of each node in the samples to be diagnosed as feature benchmark values ​​for counterfactual reasoning.

[0166] The counterfactual state inference unit is used to apply different combinations of intervention values ​​to the confounding variables in the feature baseline value based on the causal structure function, while keeping the non-confounding variables unchanged. It performs forward inference through the causal structure function to calculate the value of the state feature of each node after intervention, and generates multiple sets of counterfactual state features after intervention.

[0167] The counterfactual sample set construction unit is used to reorganize the counterfactual state features according to the dimensional organization of the multi-source tensor to construct a counterfactual sample set containing multiple counterfactual samples.

[0168] The original fault probability diagnosis unit is used to input the sample to be diagnosed into the diagnosis model for forward propagation calculation. Through graph convolutional layers and fully connected layers, it maps layer by layer and outputs the original diagnosis probability vector of each node belonging to various types of faults.

[0169] The parallel computation unit for counterfactual probabilities is used to input each sample in the counterfactual sample set into the diagnostic model for parallel computation, and obtain the corresponding counterfactual diagnostic probability vector set respectively.

[0170] The weighted fusion decontamination diagnostic unit is used to calculate the fusion weight coefficient based on the mean effect intensity of confounding variables on the voltage and current characteristics of each node in the causal effect matrix. The original diagnostic probability vector and the counterfactual diagnostic probability vector set are weighted and averaged according to the fusion weight coefficient to generate the final diagnostic result after removing the interference of confounding factors.

[0171] In one embodiment, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described intelligent fault diagnosis method for power distribution networks.

[0172] In one embodiment, this application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described intelligent fault diagnosis method for power distribution networks.

[0173] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0174] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units 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 disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0175] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in this application, and these should all be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for intelligent fault diagnosis in power distribution networks, characterized in that, Includes the following steps: S1: The acquired quasi-steady-state measurement data of the distribution network, dynamic synchronous phasor data and meteorological data are processed for spatiotemporal alignment using the dynamic time warping algorithm to construct a multi-source tensor containing three dimensions: time, node and feature. S2: Based on the multi-source tensor, the information transmission direction and intensity between any two node feature sequences are calculated using the transfer entropy algorithm, and the calculated information transmission intensity is filtered for significance to select causal relationship edges with statistical significance, generating a dynamic causal graph. S3: Based on the dynamic causal graph and the multi-source tensor, a structural equation model is introduced to quantify the causal relationship of each node. The causal structure function of each node is obtained by solving the structural equation, and the average causal effect between nodes is calculated based on the causal structure function to generate a causal effect matrix. S4: Construct a causal enhanced graph neural network based on the dynamic causal graph and the causal effect matrix. Use the dynamic causal graph as the network graph structure and embed the causal effect matrix as the causal weights for message passing into the network propagation process. Use the historical data of the multi-source tensor with fault type and location as supervision labels for training to generate a diagnostic model that integrates causal prior knowledge. S5: Based on the causal structure function, counterfactual intervention is performed on the confounding factors in the samples to be diagnosed obtained from the real-time data stream to generate a counterfactual sample set. The samples to be diagnosed and the counterfactual sample set are respectively input into the diagnostic model for probabilistic diagnosis. The fusion weight is determined by combining the effect intensity of the confounding factors in the causal effect matrix to perform weighted fusion of the diagnostic probabilities and generate the final diagnostic result after removing the interference of confounding factors.

2. The method of claim 1, wherein, S1 includes: S11: Acquire quasi-steady-state measurement data of each node in the distribution network, dynamic synchronous phasor data of each PMU deployment location, and meteorological data of each meteorological monitoring station. Perform missing value imputation and outlier removal on various types of data to generate a cleaned multi-source dataset. S12: Normalize the various types of data in the cleaned multi-source dataset, map features of different dimensions to a unified numerical range, and generate a standardized feature sequence. S13: Perform time axis alignment processing on the feature sequence, calculate the optimal alignment path between data sequences with different sampling frequencies, eliminate the phase difference caused by communication delay through path mapping, and generate time-aligned multi-source data; S14: Organize the multi-source data into a three-dimensional multi-source tensor according to three dimensions: timestamp, node number, and feature type.

3. The method of claim 1, wherein, S2 includes: S21: Perform sliding window partitioning on the voltage and current feature sequences of each node in the multi-source tensor, calculate the mutual information value between any two node feature sequences within each window, and construct an initial adjacency matrix based on the mutual information value; S22: Based on the initial adjacency matrix, perform directional calculation on the information transmission direction between each pair of nodes, determine the dominant direction of information flow by comparing the transmission entropy values ​​in different directions, and generate a directed graph structure with directional weights. S23: Perform a significance test on each directed edge in the directed graph structure, filter out weakly associated edges with a transmission entropy value lower than the preset confidence level, retain causal relationship edges with statistical significance, and generate a dynamic causal graph.

4. The method of claim 1, wherein, S3 includes: S31: Based on the set of parent nodes of each node in the dynamic causal graph, structural equation modeling is performed on the mapping relationship between the state features of each node and the features of the parent node. Basis function expansion and fitting are performed on the mapping relationship between the state features and the features of the parent node. The basis function weight coefficients are estimated by the least squares method to generate the initial structure function of each node. S32: Perform additivity decomposition on the initial structure function to separate the contributions of exogenous disturbance variables and endogenous variables, and iteratively optimize the structure function parameters by minimizing the prediction error to generate the causal structure function of each node; S33: Based on the causal structure function, apply a unit impulse perturbation to each node and perform forward propagation calculation. Apply a unit impulse perturbation to the causal structure function of each node and perform forward inference processing. By comparing the influence of the changes in the characteristics of the parent node set before and after the impulse is applied on the characteristics of the child node, traverse all node pairs to generate a causal effect matrix.

5. The method of claim 4, wherein, The formula for calculating the causal effect matrix is ​​as follows: in, Let be the strength of the modified causal effect of node j on node i. Let i be the causal structure function of node i. Let be the set of parent nodes of node i at time t. Let represent the state characteristics of node j at time t. For time-varying decay factor control parameters, Let j be the information propagation delay from node j to node i. Let the oscillation characteristic period between node j and node i be . Let t be the phase difference between node j and node i at time t, and T be the total time length.

6. The method of claim 1, wherein, S4 includes: S41: Based on the topological structure of the dynamic causal graph, construct the adjacency matrix of the graph neural network, embed the elements in the causal effect matrix as the initial causal weights for message passing into the corresponding positions of the adjacency matrix, perform causal effect normalization on the weight values ​​of each edge in the adjacency matrix, and construct a graph neural network framework that integrates causal weights. S42: Perform multi-head attention calculation on the hidden state of each node in the graph neural network framework, and splice and fuse the node's own features, neighboring node features and corresponding causal effect weights. Adaptively adjust the contribution of different causal paths through learnable attention parameters to generate causally enhanced node representations. S43: Based on the historical data of the multi-source tensor, the graph neural network framework is trained end-to-end. The fault type label and fault location label are used as supervision signals for backpropagation optimization. The network parameters are iteratively updated through the gradient descent algorithm until the loss function converges, generating a diagnostic model that integrates causal prior knowledge.

7. The method according to any one of claims 1 to 6, characterized in that, S5 includes: S51: Perform confounding factor identification processing on the samples to be diagnosed obtained from the real-time data stream, mark meteorological features and load fluctuation features as confounding variables to be intervened, and extract the state features of each node in the samples to be diagnosed as feature benchmark values ​​for counterfactual reasoning. S52: Based on the causal structure function, different combinations of intervention values ​​are applied sequentially to the confounding variables in the feature baseline values, while keeping the non-confounding variables unchanged. The causal structure function is used to perform forward inference to calculate the values ​​of the state features of each node after the intervention, generating multiple sets of counterfactual state features after the intervention. S53: Reorganize the counterfactual state features according to the dimensional organization method of the multi-source tensor to construct a counterfactual sample set containing multiple counterfactual samples; S54: Input the sample to be diagnosed into the diagnostic model for forward propagation calculation, and map it layer by layer through graph convolutional layers and fully connected layers to output the original diagnostic probability vector of each node belonging to various types of faults; S55: Input each sample in the counterfactual sample set into the diagnostic model in sequence for parallel computation to obtain the corresponding counterfactual diagnostic probability vector set; S56: Based on the mean effect intensity of the confounding variables on the voltage and current characteristics of each node in the causal effect matrix, calculate the fusion weight coefficient, and perform weighted averaging on the original diagnostic probability vector and the counterfactual diagnostic probability vector set according to the fusion weight coefficient to generate the final diagnostic result after removing the interference of confounding factors.

8. A power distribution network fault intelligent diagnosis system, characterized in that, The system includes: The multi-source data spatiotemporal alignment module is used to perform spatiotemporal alignment processing on the acquired quasi-steady-state measurement data of the distribution network, dynamic synchronous phasor data and meteorological data using a dynamic time warping algorithm, and to construct a multi-source tensor containing three dimensions: time, node and feature. The dynamic causal graph construction module is used to calculate the information transmission direction and intensity between any two node feature sequences based on the multi-source tensor using the transfer entropy algorithm, and to perform significance filtering on the calculated information transmission intensity to select causal relationship edges with statistical significance, thereby generating a dynamic causal graph. The causal effect quantification module is used to quantify the causal relationship of each node by introducing a structural equation model based on the dynamic causal graph and the multi-source tensor. It obtains the causal structure function of each node by solving the structural equation, and calculates the average causal effect between nodes based on the causal structure function to generate a causal effect matrix. The causal augmentation graph network training module is used to construct a causal augmentation graph neural network based on the dynamic causal graph and the causal effect matrix. The dynamic causal graph is used as the network graph structure and the causal effect matrix is ​​used as the causal weights for message passing and embedded in the network propagation process. The module uses the historical data of the multi-source tensor with fault type and location as supervision labels for training to generate a diagnostic model that integrates causal prior knowledge. The counterfactual diagnostic fusion module is used to perform counterfactual intervention processing on the confounding factors in the samples to be diagnosed obtained from the real-time data stream based on the causal structure function, generate a counterfactual sample set, input the samples to be diagnosed and the counterfactual sample set into the diagnostic model for probabilistic diagnosis, and combine the effect intensity of the confounding factors in the causal effect matrix to determine the fusion weight to perform weighted fusion of the diagnostic probabilities, and generate the final diagnostic result after removing the interference of confounding factors.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images