A power distribution equipment fault early warning method and system based on multi-source electric parameters

Abstract

The application relates to the technical field of fault detection, and discloses a power distribution equipment fault early warning method and system based on multi-source electric parameters, wherein the method comprises the following steps: collecting multi-source electric parameters and environmental parameter data of power distribution equipment, obtaining a standardized multi-source parameter sliding time window through synchronous correction, steady-state window screening and standardization processing; constructing a directed weighted parameter correlation graph by using the maximum information coefficient and the transfer entropy; reconstructing a normal correlation baseline by using physical bias graph attention encoding and decoding to obtain a reconstructed baseline matrix group; obtaining a correlation deviation value by combining matrix reconstruction deviation and graph structure distribution deviation; and outputting a fault early warning result based on adaptive threshold judgment and abnormal correlation positioning. According to the application, the early warning is carried out according to the changes of parameter correlation strength, causal direction and graph structure relative to the normal baseline, and an abnormal correlation edge and a score thereof are provided, so that the object to be checked can be determined.

Description

Technical Field

[0001] This invention relates to the field of fault detection technology, and in particular to a method and system for early warning of power distribution equipment faults based on multi-source electrical parameters. Background Technology

[0002] Currently, in the monitoring of the operational status of power distribution equipment such as distribution cabinets, transformer substations, and circuit breakers, it is usually necessary to simultaneously collect multiple electrical parameters, including load current, conductor temperature, dielectric loss factor, partial discharge, and ambient temperature and humidity, in order to detect latent faults such as insulation aging and joint overheating as early as possible. The industry commonly uses alarm logic based on fixed upper and lower thresholds for a single parameter, or simply uses AND and OR combinations to perform joint judgments on the out-of-limit states of multiple parameters, attempting to reduce false alarms caused by random fluctuations of a single parameter. This approach essentially assumes that each parameter is independent and that the correlation is constant, relying on manual experience to set thresholds and logical rules. It cannot perceive the dynamic migration of the coupling relationship between parameters caused by changes in operating mode and environmental conditions, and lacks the ability to distinguish between changes accompanying normal operating conditions and fault symptoms.

[0003] For example, in actual operation of power distribution equipment, a short-term surge in load current will normally cause a delayed rise in conductor temperature. Simultaneously, the dielectric loss factor will drift due to temperature effects and moisture migration, and partial discharge may also fluctuate non-fault-related due to humidity. Fixed-threshold alarm mechanisms confuse these physically correlated changes with fault characteristics such as abnormal temperature rises and continuous increases in dielectric loss caused by insulation degradation or loose connections, frequently triggering alarms during load switching, sudden changes in ambient temperature, or adjustments to operating modes. Such false alarms consume significant maintenance manpower for ineffective inspections. More seriously, a consistently high false alarm rate drastically reduces maintenance personnel's trust in the early warning system, leading to the habitual ignoring of genuine fault alarms and creating safety hazards.

[0004] Therefore, there is an urgent need for a fault early warning method that can still learn the normal dynamic correlation structure between parameters from data and detect abnormal deviations in correlation patterns in real time, even when the correlation relationships between multiple power parameters in power distribution equipment vary with operating conditions and are subject to uncertainty. This method should automatically adapt to the evolution of load states and environmental conditions, triggering an early warning only when the coupling relationship between multiple parameters breaks through the normal evolution boundary and cannot be explained by changes in operating conditions. This would significantly reduce the false alarm rate without relying on static thresholds, thereby improving the reliability and practicality of online early warning for power distribution equipment. Summary of the Invention

[0005] To address the aforementioned technical shortcomings, the present invention aims to propose a fault early warning method for power distribution equipment based on multi-source electrical parameters. This method addresses the problem of existing fault determination methods that rely on fixed rules or static models, particularly those that fail to adaptively learn and detect abnormal changes in the correlation structure between parameters under conditions where the multi-source electrical parameter relationships of power distribution equipment change with operating conditions and are uncertain, resulting in a high false alarm rate.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: The present invention provides a method for early warning of power distribution equipment faults based on multi-source electrical parameters.

[0007] The method for early warning of power distribution equipment based on multi-source electrical parameters includes: Step S10: Obtain multi-source electrical parameters and environmental parameters data of the power distribution equipment, and perform windowed preprocessing tasks based on the multi-source electrical parameters and environmental parameters data using a synchronous correction and steady-state window filtering mechanism to output a standardized multi-source parameter sliding time window; Step S20: Based on the standardized multi-source parameter sliding time window, the parameter correlation extraction task is performed using the joint directed graph construction mechanism of maximum information coefficient and transfer entropy, and a directed weighted parameter correlation graph is output. Step S30: Based on the directed weighted parametric correlation graph, a physical bias graph attention encoding and decoding mechanism is used to perform the normal correlation baseline reconstruction preprocessing task, and output the reconstructed baseline matrix group; Step S40: Based on the directed weighted parametric correlation graph and the reconstructed baseline matrix group, the correlation deviation calculation task is performed using a combination mechanism of matrix reconstruction deviation and graph structure distribution deviation, and the correlation deviation value is output. Step S50: Based on the correlation deviation value, perform the fault warning decision task using an adaptive threshold judgment and anomaly correlation location mechanism, and output the fault warning result.

[0008] Preferably, step S10, which involves acquiring multi-source electrical and environmental parameter data of the power distribution equipment, performing windowed preprocessing based on the multi-source electrical and environmental parameter data using a synchronous correction and steady-state window filtering mechanism, and outputting a standardized multi-source parameter sliding time window, specifically includes: Step S101: Synchronously collect the load current, conductor temperature, dielectric loss factor, partial discharge quantity, ambient temperature and ambient humidity of the power distribution equipment, and perform median filtering and zero drift correction on the sampling sequence of each parameter to obtain the corrected multi-source parameter sequence; Step S102: According to the preset sliding window width and preset window sliding step size, extract the sampled values ​​of each parameter in the current sliding time window from the corrected multi-source parameter sequence, calculate the median and median absolute deviation of each parameter in the current sliding time window, and perform robust standardization on the sampled values ​​of each parameter according to the median and median absolute deviation of the corresponding parameter to obtain standardized window data. Step S103: Calculate the current fluctuation range between the maximum and minimum values ​​of the load current within the current sliding time window, and calculate the ratio of the current fluctuation range to the preset reference load current; when the ratio is greater than the preset operation switching threshold, use the sampling time when the load current changes abruptly as the boundary of the operation event to segment the standardized window data, and determine the sub-windows adjacent to the current sampling time and with a duration not less than one-tenth of the preset sliding window width as steady-state sub-windows; when the ratio is not greater than the preset operation switching threshold, retain the standardized window data; determine the steady-state sub-windows or the retained standardized window data as the standardized multi-source parameter sliding time window.

[0009] Preferably, step S20, which involves performing the parameter correlation extraction task based on the standardized multi-source parameter sliding time window using a directed graph construction mechanism combining maximum information coefficient and transfer entropy, and outputting a directed weighted parameter correlation graph, specifically includes: Step S201: For any two parameters within the standardized multi-source parameter sliding time window and The sampling values ​​are assigned incremental time weights in order of their sampling time from oldest to youngest, and the parameters are estimated based on these time-weighted sampling values. With parameters The joint probability distribution is calculated, and the maximum information coefficient is calculated according to the following formula:

[0010] in, For parameters With parameters The maximum information coefficient between them For parameters Grid division, For parameters Grid division, For parameters The number of grid divisions, For parameters The number of grid divisions, Parameters for corresponding grid division With parameters mutual information, To preset the maximum number of grid divisions, This indicates taking the minimum value; the maximum information coefficient of each parameter pair is written into the current correlation matrix; Step S202: Calculate the parameters respectively Pointer parameter Transmission entropy and parameters Pointer parameter The propagation entropy, where the parameter Pointer parameter The propagation entropy is calculated according to the following formula:

[0011] in, The intrinsic parameters of the current sliding time window Pointer parameter The propagation entropy, This refers to the current moment corresponding to the sliding time window. The sampling time number within the current sliding time window. For parameters In the The sampled values ​​at each sampling time. For parameters In the The sampling time corresponding to each sampling time Rank history embedding vector, For parameters In the The sampling time corresponding to each sampling time Rank history embedding vector, To preset the historical embedding order, This represents the joint probability or conditional probability, and the summation operation covers the state combinations formed by each historical embedding vector and subsequent sampled values ​​within the current sliding time window; A permutation test at a preset significance level is performed on the difference in transfer entropy between the two directions. When the transfer entropy in one direction is significantly greater than that in the opposite direction, the one direction is determined as the causal association direction of the corresponding parameter pair, and the causal association direction is written into the current causal direction matrix. Step S203: Set each parameter as a parameter node, set the causal association direction that passes the permutation test as a directed edge between parameter nodes, set the maximum information coefficient of the parameter pair corresponding to the directed edge as the edge weight, and construct the directed weighted parameter association graph based on the parameter nodes, the directed edges and the edge weights.

[0012] Preferably, step S202, which involves performing a permutation test on the difference in transfer entropy between the two directions at a preset significance level, specifically includes: Set the preset historical embedding order to first order, and set the time delay between adjacent sampled values ​​in the historical embedding vector to a sampling time interval; The starting parameter of the causal association direction to be tested is determined as the parameter to be replaced. While keeping the sampling value arrangement order of the endpoint parameter unchanged, the sampling value arrangement order of the parameter to be replaced within the sliding time window of the standardized multi-source parameter is randomly replaced multiple times. Calculate the difference in transfer entropy between the two directions after each random permutation, and form a null hypothesis distribution based on the difference in transfer entropy corresponding to each random permutation; The transfer entropy difference obtained without random permutation is compared with the null hypothesis distribution. When the transfer entropy difference reaches a preset significance level, the causal association direction to be tested is retained. When the transfer entropy difference does not reach the preset significance level, the causal association direction to be tested is deleted.

[0013] Preferably, step S30, which involves performing a normal correlation baseline reconstruction preprocessing task based on the directed weighted parametric correlation graph using a physical bias graph attention encoding and decoding mechanism, and outputting a reconstructed baseline matrix set, specifically includes: Step S301: Extract parameter node features, the current association matrix, and the current causal direction matrix from the directed weighted parametric association graph. Input the parameter node features into the physical bias graph attention network and calculate the parameter node according to the following formula. For adjacent parameter nodes Attention coefficient:

[0014] in, For the parameter nodes within the current sliding time window For adjacent parameter nodes Attention coefficient For parameter nodes Node characteristics, For parameter nodes Node characteristics, For parameter nodes Node characteristics, It is a linear transformation matrix. For attention vectors, superscript This indicates the transpose operation. This represents vector concatenation. For parameter nodes The set of adjacent parameter nodes, The parameter node number is the number of the adjacent parameter node set. For parameter nodes With parameter nodes The physical similarity bias term between them For parameter nodes With parameter nodes The physical similarity bias term between them This represents the natural exponential function. This represents a linear unit function with leakage correction; When parameter node Load current node and parametric node When the temperature node is a conductor, the corresponding physical similarity bias term is determined according to the following formula:

[0015]

[0016] in, This is the physical similarity bias term between the load current node and the conductor temperature node within the current sliding time window. Given the current conductor temperature, This is the current predicted steady-state conductor temperature. The current ambient temperature. This is the current load current. The AC resistance of the current-carrying conductor in power distribution equipment. The equivalent thermal resistance from the current-carrying conductor of the power distribution equipment to the surrounding environment. The standard deviation of the conductor temperature prediction error under historical normal operating conditions is given; the physical similarity bias term of the remaining parameter node pairs is set to zero, and the corresponding parameter node features are weighted and aggregated according to each attention coefficient to obtain the graph embedding vector; Step S302: Perform an inner product operation on the graph embedding vectors using the decoder to obtain the reconstructed association matrix. Then, perform direction prediction on the graph embedding vectors using a bilinear mapping layer to obtain the reconstructed causal direction matrix. Finally, calculate the total reconstruction loss according to the following formula:

[0017] in, This represents the total reconstruction loss corresponding to the current sliding time window. For the current association matrix, For the reconstructed correlation matrix, Describing the Frobenius norm, For the preset directional loss weight, The set of valid directed edges in the directed weighted parametric association graph. and These are the starting node number and ending node number of a valid directed edge, respectively. For the parameter nodes in the current causal direction matrix With parameter nodes The actual direction labels between them For the parameter nodes in the reconstructed causal direction matrix With parameter nodes The predicted direction probability between them Represents the binary cross-entropy function; Step S303: Compare the total reconstruction loss with a preset normal update threshold and a preset freeze threshold respectively; when the total reconstruction loss is not greater than the preset normal update threshold, update the model parameters of the physical bias graph attention network and the decoder according to the preset baseline learning rate; when the total reconstruction loss is greater than the preset normal update threshold and less than the preset freeze threshold, decay the preset baseline learning rate according to the preset learning rate decay ratio, and update the model parameters according to the decayed learning rate; when the total reconstruction loss is not less than the preset freeze threshold, freeze the model parameters; combine the reconstruction correlation matrix and the reconstruction causal direction matrix to form the reconstruction baseline matrix group.

[0018] Preferably, step S40, which involves performing the correlation deviation calculation task based on the directed weighted parametric correlation graph and the reconstructed baseline matrix group using a combination mechanism of matrix reconstruction deviation and graph structure distribution deviation, and outputting the correlation deviation value, specifically includes: Step S401: Extract the test association matrix composed of the edge weights of each directed edge and the test causal direction matrix composed of the causal association directions of each directed edge from the directed weighted parametric association graph; extract the reconstructed association matrix and the reconstructed causal direction matrix from the reconstructed baseline matrix group; convert the predicted direction probabilities in the reconstructed causal direction matrix into reconstructed direction labels using a preset direction probability threshold; calculate the normalized matrix difference between the test association matrix and the reconstructed association matrix, and calculate the direction inconsistency ratio between the actual direction labels in the test causal direction matrix and the reconstructed direction labels; perform a weighted combination of the normalized matrix difference and the direction inconsistency ratio to obtain the matrix reconstruction bias; Step S402: Extract the weighted out-degree, weighted in-degree, clustering coefficient, and betweenness centrality of each parameter node from the directed weighted parametric association graph; perform dimensionless normalization on the extracted graph structure parameters; and form the current graph structure feature vector from the normalized graph structure parameters; obtain the historical graph structure feature vectors corresponding to a preset number of historical normal sliding time windows; form a historical normal graph structure feature distribution based on each historical graph structure feature vector; calculate the Wasserstein distance between the current graph structure feature vector and the historical normal graph structure feature distribution to obtain the graph structure distribution deviation. Step S403: Weight the matrix reconstruction deviation according to the first preset combination weight, weight the graph structure distribution deviation according to the second preset combination weight, sum the weighted matrix reconstruction deviation and the weighted graph structure distribution deviation, and determine the summation result as the correlation deviation value, wherein the sum of the first preset combination weight and the second preset combination weight is one.

[0019] Preferably, step S50, which involves performing a fault warning decision-making task based on the correlation deviation value using an adaptive threshold determination and an anomaly correlation localization mechanism, and outputting the fault warning result, specifically includes: Step S501: Obtain a historical sequence of correlation deviation values ​​formed by arranging multiple historical correlation deviation values ​​according to the sampling time, and determine the 90th percentile of the historical sequence of correlation deviation values ​​as the threshold starting point; select historical correlation deviation values ​​greater than the threshold starting point from the historical sequence of correlation deviation values, and determine the difference between the selected historical correlation deviation values ​​and the threshold starting point as the threshold exceedance; fit a generalized Pareto distribution using each threshold exceedance, and update the shape and scale parameters of the generalized Pareto distribution online according to the newly added correlation deviation values; Step S502: Calculate the cumulative sum statistic based on the continuously updated scale parameter. When the cumulative sum statistic reaches the preset drift judgment condition, redetermine the threshold starting point based on the updated historical sequence of correlation deviation values, and reinitialize the shape parameter and the scale parameter. Determine the current warning threshold based on the preset false alarm probability, the threshold starting point, the shape parameter, the scale parameter, the total number of samples in the historical sequence of correlation deviation values, and the number of samples greater than the threshold starting point. Step S503: Compare the correlation deviation value with the current warning threshold. When the correlation deviation value is greater than the current warning threshold, generate a fault warning signal. Calculate the contribution of each parameter node feature and each directed edge weight in the directed weighted parameter correlation graph to the correlation deviation value using the integral gradient method. Determine the abnormal correlation edge score based on the edge weight contribution of each directed edge and the node feature contributions of the corresponding starting and ending parameter nodes. Select a preset number of abnormal correlation edges in descending order of the abnormal correlation edge scores, and combine the fault warning signal, the selected abnormal correlation edges, and the corresponding abnormal correlation edge scores to form the fault warning result.

[0020] The present invention also provides a power distribution equipment fault early warning system based on multi-source electrical parameters, comprising: The multi-source parameter preprocessing module is used to acquire multi-source electrical parameters and environmental parameter data of power distribution equipment, and to perform windowed preprocessing tasks based on the multi-source electrical parameters and environmental parameter data using a synchronous correction and steady-state window filtering mechanism, and output a standardized multi-source parameter sliding time window. The association graph construction module is used to perform the parameter association task based on the standardized multi-source parameter sliding time window and the directed graph construction mechanism of the maximum information coefficient and the transfer entropy, and output a directed weighted parameter association graph. The baseline reconstruction module is used to perform normal correlation baseline reconstruction preprocessing tasks based on the directed weighted parametric correlation graph using a physical bias graph attention encoding and decoding mechanism, and outputs a reconstructed baseline matrix group. The deviation calculation module is used to perform the correlation deviation calculation task based on the directed weighted parametric correlation graph and the reconstructed baseline matrix group, using a combination mechanism of matrix reconstruction deviation and graph structure distribution deviation, and output the correlation deviation value. The early warning decision module is used to perform fault early warning decision tasks based on the correlation deviation value using an adaptive threshold judgment and anomaly correlation location mechanism, and output fault early warning results.

[0021] The present invention also provides a power distribution equipment fault early warning device based on multi-source electrical parameters. The power distribution equipment fault early warning device based on multi-source electrical parameters includes: a memory, a processor, and a power distribution equipment fault early warning program based on multi-source electrical parameters stored in the memory and executable on the processor. When the power distribution equipment fault early warning program based on multi-source electrical parameters is executed by the processor, it implements the above-mentioned method.

[0022] The present invention also provides a computer program product, the computer program product including a power distribution equipment fault early warning program based on multi-source electrical parameters, the power distribution equipment fault early warning program based on multi-source electrical parameters implementing the above method when executed by a processor.

[0023] The beneficial effects of this invention are as follows: 1. This invention synchronously corrects load current, conductor temperature, dielectric loss factor, partial discharge quantity, and environmental parameters, and implements robust standardization and steady-state window screening based on the median and median absolute deviation. This reduces the interference of differences in the dimensions of different parameters, sampling noise, and switching of operating states on correlation analysis, so that the standardized multi-source parameter sliding time window can characterize the parameter changes under the same stable operating conditions, providing a consistent data basis for subsequent correlation extraction.

[0024] 2. This invention combines the maximum information coefficient with the transmission entropy to construct a directed weighted parametric association graph that simultaneously represents the association strength and causal direction. It also uses the physical bias graph attention encoding and decoding mechanism to reconstruct the normal association baseline. Furthermore, it integrates the matrix reconstruction deviation and the graph structure distribution deviation to determine the association deviation value. This not only identifies the local changes in the association strength and direction of specific parameters, but also reflects the overall changes in the association network topology, thereby improving the ability to identify the precursors of power distribution equipment faults.

[0025] 3. This invention dynamically determines the early warning threshold based on the tail distribution of the historical sequence of correlation deviation, and judges the distribution drift by combining cumulative and statistical measures, so that the early warning judgment can adapt to the continuous changes in equipment operating status and environmental conditions. After the early warning is triggered, the contribution of node characteristics and directed edge weights to the correlation deviation value is analyzed by integral gradient analysis, and the fault early warning signal, abnormal correlation edge and its score are output, thereby providing a basis for operation and maintenance personnel to determine the abnormal parameter combination and the troubleshooting order. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating the first embodiment of a fault early warning method for power distribution equipment based on multi-source electrical parameters according to the present invention. Detailed Implementation

[0027] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0028] Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] Example 1: As Figure 1 The diagram shown is a flowchart of the first embodiment of a power distribution equipment fault early warning method based on multi-source electrical parameters according to the present invention. The first embodiment of the power distribution equipment fault early warning method based on multi-source electrical parameters according to the present invention is presented.

[0030] In the first embodiment, the method for early warning of power distribution equipment faults based on multi-source electrical parameters includes: Step S10: Obtain multi-source electrical parameters and environmental parameters data of the power distribution equipment, and perform windowed preprocessing tasks based on the multi-source electrical parameters and environmental parameters data using a synchronous correction and steady-state window filtering mechanism to output a standardized multi-source parameter sliding time window; Multi-source electrical and environmental parameter data for power distribution equipment refers to load current, conductor temperature, dielectric loss factor, partial discharge, ambient temperature, and ambient humidity obtained under a unified sampling clock. The acquisition end first aligns the parameter sequences according to the channel timestamps, interpolates adjacent effective values ​​for missing sampling points, and marks windows with consecutive missing values ​​exceeding a preset duration as invalid windows. Subsequently, median filtering and zero-drift correction are performed to obtain a corrected multi-source parameter sequence. After extracting the current window according to the preset sliding window width and window sliding step size, the median and absolute deviation median of each parameter are calculated. Robust standardization is achieved by dividing the deviation of the sampled value from the median by the corresponding absolute deviation median. When the absolute deviation median is zero, a regularization value determined by the channel range and acquisition resolution is used instead of the denominator. Steady-state screening is based on the ratio of the difference between the maximum and minimum load current values ​​within the current window to a preset reference load current. When this ratio is greater than a preset operation switching threshold, the moment of load current abrupt change is used as the boundary for dividing the window into operation events, and a sub-window adjacent to the current sampling moment with a duration not less than one-tenth of the window width is retained. When this ratio does not exceed the threshold, the original window is retained. The resulting standardized multi-source parameter sliding time window serves as the input for step S20.

[0031] The above processing transforms electrical, thermal, and environmental parameters with different dimensions and magnitudes to a comparable local scale, while separating transient processes such as load switching and operating mode switching from the current steady-state segment. The median and median absolute deviation are less sensitive to a few spikes than the mean and standard deviation; therefore, partial discharge pulses or communication jitter are less likely to significantly affect the standardized benchmark. Furthermore, the boundaries of operating events constrain samples within the window to belong to the same operating condition segment. In step S20, when estimating the joint distribution and directionality information of parameters, the sample scale, time reference, and operating state remain consistent, and the basis for the formation of directed edges can be traced back to valid data within the same window.

[0032] When using fixed standardization with full-time mean and standard deviation, light and heavy load data jointly determine the transformation parameters, and recent small changes may be masked by large historical fluctuations. Furthermore, without distinguishing between steady-state and transient windows, temperature lag caused by sudden load increases may be identified as a new anomalous correlation. In contrast, this approach updates robust statistics within each current window and identifies operational event boundaries based on the proportion of load current fluctuations, ensuring that subsequent correlation analysis addresses relative changes under the current operating conditions. This difference is not merely about scaling the data; it simultaneously limits the sample range and temporal consistency involved in correlation estimation, reducing the interference of operating condition switching on causal direction judgment from the data entry point.

[0033] For example, a 10kV power distribution device has a load current of approximately 320 amps during the daytime peak period, which drops to 80 amps during the nighttime off-peak period, while the conductor temperature slowly decreases due to thermal inertia. If the current window spans the moment of load decline, the current has already entered the off-peak period, but the temperature still retains the characteristics of the peak period. Directly calculating the correlation would underestimate the coupling between the two under steady-state conditions. According to this scheme, when the current fluctuation ratio exceeds the operating switching threshold, the window is segmented at the moment of abrupt change. The system selects the off-peak sub-window adjacent to the current sampling time and satisfies the minimum duration, and then robustly normalizes each parameter. The resulting window only describes the off-peak operating condition and can be directly used in step S20 to construct the current directed weighted parameter correlation graph.

[0034] Step S20: Based on the standardized multi-source parameter sliding time window, the parameter correlation extraction task is performed using the joint directed graph construction mechanism of maximum information coefficient and transfer entropy, and a directed weighted parameter correlation graph is output. The standardized multi-source parameter sliding time window is a dataset output from step S10, with sampling times aligned across channels and filtered for steady-state performance. For any two parameters, increasing time weights are assigned based on sampling times, from earliest to latest, ensuring that recent samples contribute more to the joint probability distribution than earlier samples. Candidate grids are enumerated within a preset upper limit for the number of grid divisions, and mutual information is calculated for each grid. This information is then normalized using the binary logarithm of the smaller grid division number, and the maximum value of the normalized result is taken as the maximum information coefficient. The maximum information coefficient of each parameter pair is then written into the current association matrix. For direction determination, the first-order historical transfer entropy of each direction is calculated, with a historical time delay of one sampling interval. The sample order of the starting parameter is randomly permuted multiple times to form a null hypothesis distribution of the transfer entropy difference. Only when the original difference reaches a preset significance level is the corresponding direction retained. Parameters are treated as nodes, retained directions as directed edges, and the maximum information coefficient as edge weights, together forming a directed weighted parameter association graph. If the number of effective samples is insufficient or the permutation test results are unstable, no new direction is forced; instead, the parameter pair is marked as pending confirmation and awaits subsequent effective windows.

[0035] The maximum information coefficient describes the strength of the linear or nonlinear correlation between parameters, while the transfer entropy and its permutation test determine whether information changes propagate along a specific direction. Both determine the edge weights and edge directions for the same parameter pair, respectively. The resulting graph retains the correlation between load current, temperature, dielectric loss, and partial discharge while excluding accidental directions that do not meet the significance requirement. After receiving the graph in step S30, encoding and decoding can be performed using node features, the current correlation matrix, and the current causal direction matrix, respectively, avoiding mixing undirected correlations and directed causal hints into the same input.

[0036] When using only Pearson correlation coefficients or ordinary mutual information, the output is usually a symmetric matrix. Synchronous changes in ambient temperature and load current due to seasonality or electricity consumption patterns may be mistakenly interpreted as a direct driving relationship. Furthermore, when processing the entire window with fixed weights, earlier samples can weaken recent correlation changes. In contrast, this scheme preserves nonlinear relationships with the maximum information coefficient, filters out accidental directions using bidirectional transitive entropy differences and permutation tests, and increases the weight of recent samples in probability estimation. The resulting edges must simultaneously possess identifiable correlation strength and statistically acceptable directional evidence, making them more suitable as graph structure inputs for reconstructing normal correlation baselines.

[0037] Step S30: Based on the directed weighted parametric correlation graph, a physical bias graph attention encoding and decoding mechanism is used to perform the normal correlation baseline reconstruction preprocessing task, and output the reconstructed baseline matrix group; A directed weighted parametric correlation graph (CWC) is a current window graph composed of parametric nodes, directed edges that pass the significance test, and edge weights with the maximum information coefficient. During processing, parametric node features, the current correlation matrix, and the current causal direction matrix are extracted from the graph, and the node features are input into a physical bias graph attention network. For each node and its adjacent nodes, a linear transformation and concatenation of their features are performed first, followed by the addition of the corresponding physical similarity bias. Attention coefficients are obtained after processing with a leakage-corrected linear unit and normalization. The physical biases of the load current node and the conductor temperature node are determined based on the deviation of the measured conductor temperature from the predicted steady-state conductor temperature. The predicted steady-state conductor temperature is obtained by adding the product of the current ambient temperature, the square of the load current, the AC resistance of the current-carrying conductor, and the equivalent thermal resistance. The AC resistance can be corrected by equipment parameters and operating temperature; the equivalent thermal resistance is identified through temperature rise data during normal operation; the temperature prediction error scale is obtained from historical normal residual statistics; and the physical biases of other node pairs are set to zero. The decoder performs an inner product of the graph embedding vectors to obtain the reconstructed association matrix, and then obtains the reconstructed causal direction matrix through a bilinear mapping. The total reconstruction loss is formed by weighting the sum of squared differences in the association matrices with the binary cross-entropy of the direction predictions. The learning rate is used to update the matrix when the total reconstruction loss is not greater than the normal update threshold; it is used to update the matrix with a decaying learning rate when it is between the normal update threshold and the freezing threshold; and updates stop when the freezing threshold is reached. The two reconstruction matrices together constitute the reconstruction baseline matrix set.

[0038] The physical similarity bias enables the attention network to verify whether the relationship between load current and conductor temperature conforms to an interpretable thermal equilibrium when learning normal data distribution. The decoding end simultaneously reconstructs the correlation strength and causal direction, ensuring that the normal baseline includes not only edge weights but also directional probabilities. The hierarchical update rule uses the total reconstructed loss as the basis for whether the model continues to absorb the current window: normal updates occur when the loss is small, the update magnitude is reduced when in the transition range, and existing parameters are maintained when the loss reaches the freezing threshold. This avoids the abnormal window rapidly altering the normal baseline and provides a reference matrix within the same window for step S40 to calculate the intensity bias and direction bias respectively.

[0039] Conventional autoencoders update parameters solely based on data reconstruction errors. Even if a relationship between load current and conductor temperature deviates from thermal law, it may be gradually absorbed with continuous input. A fixed learning rate can also cause slowly developing fault characteristics to be treated as new normal patterns. In contrast, this scheme incorporates a physical bias determined by ambient temperature, load current, AC resistance, and equivalent thermal resistance during the attention calculation phase, and decays or freezes the update as the total reconstruction loss increases. Therefore, the model retains the data-driven capability without prior parameter pairs while imposing consistency constraints on node pairs with clear physical relationships, such as current and temperature.

[0040] For example, when the bolts of the terminal joints of power distribution equipment loosen, the contact resistance gradually increases, and the measured conductor temperature will remain higher than the steady-state prediction value obtained based on the load current and ambient temperature. At this time, the physical similarity between the load current and the conductor temperature decreases, the graph attention allocation and decoding results begin to deviate from the normal state, and the total reconstruction loss increases accordingly. Before the loss reaches the freezing threshold, the model is updated slowly with a decaying learning rate; after the loss reaches the freezing threshold, the absorption of this window stops. The output reconstruction correlation matrix and reconstruction causal direction matrix still maintain the normal baseline characteristics, and step S40 can identify the difference between the current graph and the normal baseline based on this.

[0041] Step S40: Based on the directed weighted parametric correlation graph and the reconstructed baseline matrix group, the correlation deviation calculation task is performed using a combination mechanism of matrix reconstruction deviation and graph structure distribution deviation, and the correlation deviation value is output. The directed weighted parametric association graph provides a test association matrix composed of current edge weights and a test causal direction matrix composed of current causal directions. The reconstructed baseline matrix set provides the corresponding reconstructed association matrix and reconstructed causal direction matrix. First, the reconstructed causal direction probabilities are converted into direction labels according to a preset direction probability threshold. Then, the norm difference between the test association matrix and the reconstructed association matrix is ​​normalized relative to the scale of the test matrix, and the proportion of edges with inconsistent test and reconstructed direction labels is statistically analyzed. The two are combined according to preset weights to form the matrix reconstruction deviation. Subsequently, the weighted out-degree, weighted in-degree, clustering coefficient, and betweenness centrality of each node are extracted from the current graph. Graph structure parameters of different dimensions are normalized to form the current graph structure feature vector. A normal graph structure feature distribution is formed using a preset number of historical normal windows. The Wasserstein distance between the current vector and this distribution is calculated to obtain the graph structure distribution deviation. Finally, the association deviation value is obtained by summing the weights of the two combined terms that sum to one. When the scale of the matrix to be tested is zero, normalization is performed using a preset regularization lower bound. When the historical normal window is insufficient, the most recent effective distribution is used and the baseline distribution update is paused.

[0042] Matrix reconstruction bias focuses on whether specific edge weights and causal directions deviate from the corresponding baseline, while graph structure distribution bias focuses on whether the overall topology, such as weighted degree, clustering characteristics, and betweenness centrality, deviates from the normal distribution. These two types of biases complement each other within the same associated bias value: when local edges begin to change, the matrix terms can first reflect the shift in the corresponding relationships; when multiple edges collectively change the network organization, the distribution terms can reflect the cumulative changes at the topological level. Step S50 only needs to receive a dimensionless continuous bias value to perform unified tail distribution modeling for different windows, while still tracing the source of deviation along the current matrix and graph structure features.

[0043] When using only matrix reconstruction error as the outlier score, the overall edge weights may be affected by normal load changes, causing the score to rise, but this cannot determine whether the graph topology is still in a normal distribution. When using only graph structure distance, the direction reversal of a few key edges may be diluted by the overall statistics. In contrast, this scheme calculates the corresponding difference between the current matrix and the reconstructed matrix, the proportion of inconsistencies in direction, and the graph structure distribution distance separately and then combines them, so that both local correlation changes and overall topological changes have clear points of reference. Limiting the sum of the two weights to one also facilitates calibration based on false positives and false negatives on normal validation data without changing the interpretable range of correlation deviation.

[0044] For example, after the start of the rainy season, the ambient humidity continues to rise, and the edge weights between the dielectric loss factor and the partial discharge quantity initially increase slowly, while other edges in the network remain relatively stable. Initially, the difference between the measured correlation matrix and the reconstructed correlation matrix increases first, and the graph structure distribution distance changes little. As the humidity level increases, the weighting degree, clustering coefficient, and betweenness centrality of the ambient humidity nodes and the partial discharge nodes change simultaneously, and the graph structure distribution distance increases accordingly. After the combination of these two types of deviations, the correlation deviation value shows a continuous upward trend and serves as the input for step S50; maintenance personnel can also correlate the matrix differences with the specific parameter edges that are affected.

[0045] Step S50: Based on the correlation deviation value, perform the fault warning decision task using an adaptive threshold judgment and anomaly correlation location mechanism, and output the fault warning result.

[0046] The correlation deviation value is a weighted measure of the current window anomaly obtained in step S40 by weighting the matrix reconstruction deviation and the graph structure distribution deviation. The system saves historical deviation values ​​at each sampling time, using the 90th percentile of the historical sequence as the threshold starting point. Historical values ​​exceeding this starting point are selected, and the portion of each historical value exceeding the starting point is used as the threshold exceedance. These exceedances are used to fit a generalized Pareto distribution, and the shape and scale parameters are updated online with new deviation values. The cumulative sum statistic is calculated for the continuous scale parameters. When the preset drift judgment condition is reached, the threshold starting point is redefined based on the updated historical sequence, and the distribution parameters are reinitialized. The current warning threshold is jointly determined by the preset false alarm probability, threshold starting point, shape parameters, scale parameters, total number of historical samples, and number of exceedance samples. When the current deviation value exceeds this threshold, a fault warning signal is generated, and the contribution of node features and edge weights to the current deviation is calculated using the integral gradient method. The edge weight contribution of each edge, together with the feature contributions of its starting and ending nodes, forms an abnormal correlation edge score. A preset number of edges are selected in descending order of score, and combined with the warning signal and corresponding scores to form the fault warning result. If the number of samples exceeding the limit is insufficient for fitting or the parameter update does not converge, continue using the most recent effective threshold and mark the current result as pending update.

[0047] The adaptive threshold uses the tail of the historical deviation sequence as the modeling object, and the threshold starting point and distribution parameters are updated with the accumulation of normal data; the accumulation and drift judgment avoids continuing to use old parameters after the tail distribution has changed. After the warning is triggered, the integral gradient decomposes the comprehensive deviation into node features and directed edge weights, so the abnormal correlation edge score reflects the contribution of the edge itself and the parameters at both ends to the result. The output not only indicates whether a warning has been issued, but also gives the parameter correlations that need to be checked first and their ranking criteria, so that steps S10 to S50 form a closed loop from original sampling, correlation mapping, baseline reconstruction, deviation calculation to warning location.

[0048] Fixed empirical thresholds are typically calibrated during the initial commissioning phase. However, as equipment ages, seasons change, or load structures alter, the deviation from the baseline slowly drifts, and the original threshold may continue to generate false alarms or gradually lose sensitivity. Systems that only output a comprehensive alarm value still require manual inspection of all parameters. In contrast, this solution determines the current threshold based on a preset false alarm probability distribution from the tail end, and recalibrates the threshold starting point when the scale parameter continues to drift, providing clear data support for threshold changes. Anomaly localization then assigns the early warning results to specific edges in a directed weighted parametric correlation graph, allowing maintenance personnel to verify relevant sensor channels, insulation components, or current-carrying connection points.

[0049] For example, a power distribution system that has been in operation for many years may experience a slow increase in the dielectric loss factor benchmark due to natural aging, but the normal correlation structure between various parameters remains relatively stable. As the tail of the historical deviation changes slowly, the scale parameter of the generalized Pareto distribution is continuously updated; after the cumulative sum and statistics reach the drift condition, the system uses the new historical sequence to redetermine the threshold starting point, thereby avoiding the use of the initial threshold during commissioning. Subsequently, if a joint experiences localized overheating, and the current correlation deviation value exceeds the updated warning threshold, the system generates a warning signal. If the integral gradient analysis shows that the two edges—partial discharge pointing to conductor temperature and load current pointing to conductor temperature—contribute the most, the fault warning result simultaneously lists these two abnormal correlation edges and their corresponding scores, allowing maintenance personnel to prioritize checking the joint insulation and connection status.

[0050] Example 2: Furthermore, the present invention provides a power distribution equipment fault early warning system based on multi-source electrical parameters, employing a power distribution equipment fault early warning method based on multi-source electrical parameters as described in the above embodiments, which can solve the technical problem of power distribution equipment fault early warning based on multi-source electrical parameters. The beneficial effects of the power distribution equipment fault early warning system based on multi-source electrical parameters provided by the present invention are the same as those of the power distribution equipment fault early warning method based on multi-source electrical parameters provided in the above embodiments, and other technical features of the power distribution equipment fault early warning system based on multi-source electrical parameters are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0051] Example 3: This invention provides a power distribution equipment fault early warning device based on multi-source electrical parameters. The device includes at least one processor and a memory communicatively connected to the processor. The memory stores instructions executable by the processor, which are then executed to enable the processor to perform the power distribution equipment fault early warning method based on multi-source electrical parameters described in Example 1. This invention's power distribution equipment fault early warning device based on multi-source electrical parameters can include, but is not limited to, mobile terminals such as mobile phones, laptops, digital radio receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), and vehicle terminals (e.g., vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. This power distribution equipment fault early warning device based on multi-source electrical parameters is merely an example and should not limit the functionality or scope of the invention. A fault warning device for power distribution equipment based on multi-source electrical parameters may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) that can perform various appropriate actions and processes according to a program stored in a read-only memory or a program loaded from a storage device into a random access memory. The random access memory also stores various programs and data required for the operation of the fault warning device. The processing unit, the read-only memory, and the random access memory are interconnected via a bus. An I / O interface is also connected to the bus. Typically, the following systems can be connected to the I / O interface: input devices including touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices including liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices including magnetic tapes, hard disks, etc.; and communication devices. The communication device allows the fault warning device for power distribution equipment based on multi-source electrical parameters to communicate wirelessly or wiredly with other devices to exchange data. Although a fault warning device for power distribution equipment based on multi-source electrical parameters with various systems has been described, it should be understood that it is not required to implement or possess all the systems described. It can be implemented alternatively or have more or fewer systems.

[0052] Example 4: This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described method for early warning of power distribution equipment faults based on multi-source electrical parameters. The computer program product provided by this invention can solve the technical problem of early warning of power distribution equipment faults based on multi-source electrical parameters. Compared with the prior art, the beneficial effects of the computer program product provided by this invention are the same as those of the method for early warning of power distribution equipment faults based on multi-source electrical parameters provided in the above embodiments, and will not be repeated here.

[0053] In particular, according to the embodiments disclosed in this invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from a storage device, or installed from a read-only memory. When the computer program is executed by a processing device, it performs the functions defined in the methods of the embodiments disclosed in this invention.

[0054] It should be understood that the various parts disclosed in this invention can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics may be combined in any suitable manner in one or more embodiments or examples.

[0055] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the present invention and its equivalents, the present invention also intends to include these modifications and variations.

Claims

1. A method for early warning of power distribution equipment faults based on multi-source electrical parameters, characterized in that, The method includes: Step S10: Obtain multi-source electrical parameters and environmental parameters data of the power distribution equipment, and perform windowed preprocessing tasks based on the multi-source electrical parameters and environmental parameters data using a synchronous correction and steady-state window filtering mechanism to output a standardized multi-source parameter sliding time window; Step S20: Based on the standardized multi-source parameter sliding time window, the parameter correlation extraction task is performed using the joint directed graph construction mechanism of maximum information coefficient and transfer entropy, and a directed weighted parameter correlation graph is output. Step S30: Based on the directed weighted parametric correlation graph, a physical bias graph attention encoding and decoding mechanism is used to perform the normal correlation baseline reconstruction preprocessing task, and output the reconstructed baseline matrix group; Step S40: Based on the directed weighted parametric correlation graph and the reconstructed baseline matrix group, the correlation deviation calculation task is performed using a combination mechanism of matrix reconstruction deviation and graph structure distribution deviation, and the correlation deviation value is output. Step S50: Based on the correlation deviation value, perform the fault warning decision task using an adaptive threshold judgment and anomaly correlation location mechanism, and output the fault warning result.

2. The method for early warning of power distribution equipment faults based on multi-source electrical parameters as described in claim 1, characterized in that, In step S10, the steps of acquiring multi-source electrical parameters and environmental parameters of the power distribution equipment, performing windowed preprocessing based on the multi-source electrical parameters and environmental parameters using a synchronous correction and steady-state window filtering mechanism, and outputting a standardized multi-source parameter sliding time window specifically include: Step S101: Synchronously collect the load current, conductor temperature, dielectric loss factor, partial discharge quantity, ambient temperature and ambient humidity of the power distribution equipment, and perform median filtering and zero drift correction on the sampling sequence of each parameter to obtain the corrected multi-source parameter sequence; Step S102: According to the preset sliding window width and preset window sliding step size, extract the sampled values ​​of each parameter in the current sliding time window from the corrected multi-source parameter sequence, calculate the median and median absolute deviation of each parameter in the current sliding time window, and perform robust standardization on the sampled values ​​of each parameter according to the median and median absolute deviation of the corresponding parameter to obtain standardized window data. Step S103: Calculate the current fluctuation range between the maximum and minimum values ​​of the load current within the current sliding time window, and calculate the ratio of the current fluctuation range to the preset reference load current; when the ratio is greater than the preset operation switching threshold, use the sampling time when the load current changes abruptly as the boundary of the operation event to segment the standardized window data, and determine the sub-windows adjacent to the current sampling time and with a duration not less than one-tenth of the preset sliding window width as steady-state sub-windows; when the ratio is not greater than the preset operation switching threshold, retain the standardized window data; determine the steady-state sub-windows or the retained standardized window data as the standardized multi-source parameter sliding time window.

3. The method for early warning of power distribution equipment faults based on multi-source electrical parameters as described in claim 1, characterized in that, Step S20, which involves extracting parameter relationships using a joint directed graph construction mechanism of maximum information coefficient and transfer entropy based on the standardized multi-source parameter sliding time window, and outputting a directed weighted parameter relationship graph, specifically includes: Step S201: For any two parameters within the standardized multi-source parameter sliding time window and The sampling values ​​are assigned incremental time weights in order of their sampling time from oldest to youngest, and the parameters are estimated based on these time-weighted sampling values. With parameters The joint probability distribution is calculated, and the maximum information coefficient is calculated according to the following formula: in, For parameters With parameters The maximum information coefficient between them For parameters Grid division, For parameters Grid division, For parameters The number of grid divisions, For parameters The number of grid divisions, Parameters for corresponding grid division With parameters mutual information, To preset the maximum number of grid divisions, This indicates taking the minimum value; the maximum information coefficient of each parameter pair is written into the current correlation matrix; Step S202: Calculate the parameters respectively Pointer parameter Transmission entropy and parameters Pointer parameter The propagation entropy, where the parameter Pointer parameter The propagation entropy is calculated according to the following formula: in, The intrinsic parameters of the current sliding time window Pointer parameter The propagation entropy, This refers to the current moment corresponding to the sliding time window. The sampling time number within the current sliding time window. For parameters In the The sampled values ​​at each sampling time. For parameters In the The sampling time corresponding to each sampling time Rank history embedding vector, For parameters In the The sampling time corresponding to each sampling time Rank history embedding vector, To preset the historical embedding order, This represents the joint probability or conditional probability, and the summation operation covers the state combinations formed by each historical embedding vector and subsequent sampled values ​​within the current sliding time window; A permutation test at a preset significance level is performed on the difference in transfer entropy between the two directions. When the transfer entropy in one direction is significantly greater than that in the opposite direction, the one direction is determined as the causal association direction of the corresponding parameter pair, and the causal association direction is written into the current causal direction matrix. Step S203: Set each parameter as a parameter node, set the causal association direction that passes the permutation test as a directed edge between parameter nodes, set the maximum information coefficient of the parameter pair corresponding to the directed edge as the edge weight, and construct the directed weighted parameter association graph based on the parameter nodes, the directed edges and the edge weights.

4. The method for early warning of power distribution equipment faults based on multi-source electrical parameters as described in claim 3, characterized in that, Step S202, which involves performing a permutation test on the difference in transmission entropy between the two directions at a preset significance level, specifically includes: Set the preset historical embedding order to first order, and set the time delay between adjacent sampled values ​​in the historical embedding vector to a sampling time interval; The starting parameter of the causal association direction to be tested is determined as the parameter to be replaced. While keeping the sampling value arrangement order of the endpoint parameter unchanged, the sampling value arrangement order of the parameter to be replaced within the sliding time window of the standardized multi-source parameter is randomly replaced multiple times. Calculate the difference in transfer entropy between the two directions after each random permutation, and form a null hypothesis distribution based on the difference in transfer entropy corresponding to each random permutation; The transfer entropy difference obtained without random permutation is compared with the null hypothesis distribution. When the transfer entropy difference reaches a preset significance level, the causal association direction to be tested is retained. When the transfer entropy difference does not reach the preset significance level, the causal association direction to be tested is deleted.

5. The method for early warning of power distribution equipment faults based on multi-source electrical parameters as described in claim 3, characterized in that, Step S30, which involves performing a normal correlation baseline reconstruction preprocessing task based on the directed weighted parametric correlation graph using a physical bias graph attention encoding and decoding mechanism, and outputting a reconstructed baseline matrix set, specifically includes: Step S301: Extract parameter node features, the current association matrix, and the current causal direction matrix from the directed weighted parametric association graph. Input the parameter node features into the physical bias graph attention network and calculate the parameter node according to the following formula. For adjacent parameter nodes Attention coefficient: in, For the parameter nodes within the current sliding time window For adjacent parameter nodes Attention coefficient For parameter nodes Node characteristics, For parameter nodes Node characteristics, For parameter nodes Node characteristics, It is a linear transformation matrix. For attention vectors, superscript This indicates the transpose operation. This represents vector concatenation. For parameter nodes The set of adjacent parameter nodes, The parameter node number is the number of the adjacent parameter node set. For parameter nodes With parameter nodes The physical similarity bias term between them For parameter nodes With parameter nodes The physical similarity bias term between them This represents the natural exponential function. This represents a linear unit function with leakage correction; When parameter node Load current node and parametric node When the temperature node is a conductor, the corresponding physical similarity bias term is determined according to the following formula: in, This is the physical similarity bias term between the load current node and the conductor temperature node within the current sliding time window. The current conductor temperature, This is the current predicted steady-state conductor temperature. The current ambient temperature. This is the current load current. The AC resistance of the current-carrying conductor in power distribution equipment. The equivalent thermal resistance from the current-carrying conductor of the power distribution equipment to the surrounding environment. The standard deviation of the conductor temperature prediction error under historical normal operating conditions is given; the physical similarity bias term of the remaining parameter node pairs is set to zero, and the corresponding parameter node features are weighted and aggregated according to each attention coefficient to obtain the graph embedding vector; Step S302: Perform an inner product operation on the graph embedding vectors using the decoder to obtain the reconstructed association matrix. Then, perform direction prediction on the graph embedding vectors using a bilinear mapping layer to obtain the reconstructed causal direction matrix. Finally, calculate the total reconstruction loss according to the following formula: in, This represents the total reconstruction loss corresponding to the current sliding time window. For the current association matrix, For the reconstructed correlation matrix, Describing the Frobenius norm, For the preset directional loss weight, The set of valid directed edges in the directed weighted parametric association graph. and These are the starting node number and ending node number of a valid directed edge, respectively. For the parameter nodes in the current causal direction matrix With parameter nodes The actual direction labels between them For the parameter nodes in the reconstructed causal direction matrix With parameter nodes The predicted direction probability between them Represents the binary cross-entropy function; Step S303: Compare the total reconstruction loss with a preset normal update threshold and a preset freeze threshold respectively; when the total reconstruction loss is not greater than the preset normal update threshold, update the model parameters of the physical bias graph attention network and the decoder according to the preset baseline learning rate; when the total reconstruction loss is greater than the preset normal update threshold and less than the preset freeze threshold, decay the preset baseline learning rate according to the preset learning rate decay ratio, and update the model parameters according to the decayed learning rate; when the total reconstruction loss is not less than the preset freeze threshold, freeze the model parameters; combine the reconstruction correlation matrix and the reconstruction causal direction matrix to form the reconstruction baseline matrix group.

6. The method for early warning of power distribution equipment faults based on multi-source electrical parameters as described in claim 1, characterized in that, Step S40, which involves calculating the correlation deviation based on the directed weighted parametric correlation graph and the reconstructed baseline matrix group using a combination mechanism of matrix reconstruction deviation and graph structure distribution deviation, and outputting the correlation deviation value, specifically includes: Step S401: Extract the test association matrix composed of the edge weights of each directed edge and the test causal direction matrix composed of the causal association directions of each directed edge from the directed weighted parametric association graph; extract the reconstructed association matrix and the reconstructed causal direction matrix from the reconstructed baseline matrix group; convert the predicted direction probabilities in the reconstructed causal direction matrix into reconstructed direction labels using a preset direction probability threshold; calculate the normalized matrix difference between the test association matrix and the reconstructed association matrix, and calculate the direction inconsistency ratio between the actual direction labels in the test causal direction matrix and the reconstructed direction labels; perform a weighted combination of the normalized matrix difference and the direction inconsistency ratio to obtain the matrix reconstruction bias; Step S402: Extract the weighted out-degree, weighted in-degree, clustering coefficient, and betweenness centrality of each parameter node from the directed weighted parametric association graph; perform dimensionless normalization on the extracted graph structure parameters; and form the current graph structure feature vector from the normalized graph structure parameters; obtain the historical graph structure feature vectors corresponding to a preset number of historical normal sliding time windows; form a historical normal graph structure feature distribution based on each historical graph structure feature vector; calculate the Wasserstein distance between the current graph structure feature vector and the historical normal graph structure feature distribution to obtain the graph structure distribution deviation. Step S403: Weight the matrix reconstruction deviation according to the first preset combination weight, weight the graph structure distribution deviation according to the second preset combination weight, sum the weighted matrix reconstruction deviation and the weighted graph structure distribution deviation, and determine the summation result as the correlation deviation value, wherein the sum of the first preset combination weight and the second preset combination weight is one.

7. The method for early warning of power distribution equipment faults based on multi-source electrical parameters as described in claim 1, characterized in that, Step S50, which involves performing a fault warning decision-making task based on the correlation deviation value using an adaptive threshold determination and anomaly correlation localization mechanism, and outputting the fault warning result, specifically includes: Step S501: Obtain a historical sequence of correlation deviation values ​​formed by arranging multiple historical correlation deviation values ​​according to the sampling time, and determine the 90th percentile of the historical sequence of correlation deviation values ​​as the threshold starting point; select historical correlation deviation values ​​greater than the threshold starting point from the historical sequence of correlation deviation values, and determine the difference between the selected historical correlation deviation values ​​and the threshold starting point as the threshold exceedance; fit a generalized Pareto distribution using each threshold exceedance, and update the shape and scale parameters of the generalized Pareto distribution online according to the newly added correlation deviation values; Step S502: Calculate the cumulative sum statistic based on the continuously updated scale parameter. When the cumulative sum statistic reaches the preset drift judgment condition, redetermine the threshold starting point based on the updated historical sequence of correlation deviation values, and reinitialize the shape parameter and the scale parameter. Determine the current warning threshold based on the preset false alarm probability, the threshold starting point, the shape parameter, the scale parameter, the total number of samples in the historical sequence of correlation deviation values, and the number of samples greater than the threshold starting point. Step S503: Compare the correlation deviation value with the current warning threshold. When the correlation deviation value is greater than the current warning threshold, generate a fault warning signal. Calculate the contribution of each parameter node feature and each directed edge weight in the directed weighted parameter correlation graph to the correlation deviation value using the integral gradient method. Determine the abnormal correlation edge score based on the edge weight contribution of each directed edge and the node feature contributions of the corresponding starting and ending parameter nodes. Select a preset number of abnormal correlation edges in descending order of the abnormal correlation edge scores, and combine the fault warning signal, the selected abnormal correlation edges, and the corresponding abnormal correlation edge scores to form the fault warning result.

8. A fault early warning system for power distribution equipment based on multi-source electrical parameters, applied to the fault early warning method for power distribution equipment based on multi-source electrical parameters as described in any one of claims 1 to 7, characterized in that, The system includes: The multi-source parameter preprocessing module is used to acquire multi-source electrical parameters and environmental parameter data of power distribution equipment, and to perform windowed preprocessing tasks based on the multi-source electrical parameters and environmental parameter data using a synchronous correction and steady-state window filtering mechanism, and output a standardized multi-source parameter sliding time window. The association graph construction module is used to perform the parameter association task based on the standardized multi-source parameter sliding time window and the directed graph construction mechanism of the maximum information coefficient and the transfer entropy, and output a directed weighted parameter association graph. The baseline reconstruction module is used to perform normal correlation baseline reconstruction preprocessing tasks based on the directed weighted parametric correlation graph using a physical bias graph attention encoding and decoding mechanism, and outputs a reconstructed baseline matrix group. The deviation calculation module is used to perform the correlation deviation calculation task based on the directed weighted parametric correlation graph and the reconstructed baseline matrix group, using a combination mechanism of matrix reconstruction deviation and graph structure distribution deviation, and output the correlation deviation value. The early warning decision module is used to perform fault early warning decision tasks based on the correlation deviation value using an adaptive threshold judgment and anomaly correlation location mechanism, and output fault early warning results.

9. A fault early warning device for power distribution equipment based on multi-source electrical parameters, characterized in that, The power distribution equipment fault early warning device based on multi-source electrical parameters includes: a memory, a processor, and a power distribution equipment fault early warning program based on multi-source electrical parameters stored in the memory and executable on the processor. When the power distribution equipment fault early warning program based on multi-source electrical parameters is executed by the processor, it implements a power distribution equipment fault early warning method based on multi-source electrical parameters as described in any one of claims 1 to 7.

10. A computer program product, characterized in that, The computer program product includes a power distribution equipment fault early warning program based on multi-source electrical parameters. When the power distribution equipment fault early warning program based on multi-source electrical parameters is executed by the processor, it implements a power distribution equipment fault early warning method based on any one of claims 1 to 7.

Images