Fault positioning method of power distribution network and related equipment

Abstract

The embodiment of the invention discloses a power distribution network fault positioning method and related equipment, and the method comprises the steps: generating a standardized feature matrix based on multi-dimensional data, such as electrical quantity, time sequence features, load features and equipment states, and carrying out the weighted fusion and information fusion analysis to obtain a fusion feature vector; the fusion feature vector can fully reflect key criterion information in a fault scene; and combining prediction result probability distribution, fusing the feature vector and a first fault positioning prediction result to calculate a positioning error, and optimizing a fault positioning model through the positioning error to form a closed-loop optimization mechanism, so that the fault positioning model has online learning and adaptive evolution capabilities, and the fault positioning accuracy is improved. The robustness and the accuracy of the fault positioning model in a complex scene are remarkably improved, so that an accurate fault positioning result is obtained.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power system control, and in particular to a fault positioning method for a power distribution network and related equipment. BACKGROUND

[0002] In recent years, with the continuous expansion of active power distribution networks and the continuous improvement of distributed energy penetration, cloud-edge collaborative fault positioning technology has gradually become a research hotspot. Existing technologies mostly adopt a combination of multi-source data fusion and intelligent algorithms, collect real-time operation data on the edge side, and build a fault positioning model on the cloud side to preliminarily determine the fault area, thereby significantly improving the timeliness of fault handling. The introduction of simulation verification based on digital twinning and dynamic weight adjustment further enhances the adaptability of the positioning process.

[0003] However, the existing methods still have obvious deficiencies in dynamic error evaluation and model iteration optimization. Most schemes rely on fixed thresholds and static correction of historical data, which cannot adapt to the continuous changes in the operation state of the power distribution network, resulting in a significant decline in positioning accuracy in complex fault scenarios, and making it impossible to obtain accurate fault positioning results. SUMMARY

[0004] Therefore, the present application provides a fault positioning method for a power distribution network and related equipment.

[0005] The specific technical scheme of the first embodiment of the present application is as follows: a fault positioning method for a power distribution network, the method comprising: collecting a first data set of the power distribution network in a current fault state on the edge side; the data set comprising electrical quantity data, time series feature data, load characteristic data and device state data; inputting the first data set into a preset fault positioning model to obtain a first fault positioning prediction result of the first data set output by the preset fault positioning model; obtaining a standardized feature matrix corresponding to the first data set, performing weighted summation and information fusion analysis on the standardized feature matrix to obtain a fusion feature vector corresponding to the first data set; obtaining a prediction result probability distribution of the preset fault positioning model, and performing bias quantification evaluation according to the fusion feature vector, the first fault positioning prediction result and the prediction result probability distribution to obtain a positioning error of the preset fault positioning model; adjusting the parameters and weights of the preset fault positioning model according to the positioning error to obtain an optimized fault positioning model; and optimizing the first fault positioning prediction result using the optimized fault positioning model to obtain an optimized fault positioning result.

[0006] Preferably, the obtaining the prediction result probability distribution of the preset fault location model comprises: obtaining a plurality of second data sets of the power distribution network in different historical fault states, wherein one second data set corresponds to one historical fault state; inputting the second data set into the preset fault location model to obtain a second fault location prediction result of each second data set output by the preset fault location model; and obtaining the prediction result probability distribution of the preset fault location model according to the second fault location prediction result and an actual fault location result of different historical fault states.

[0007] Preferably, the bias quantification evaluation according to the fusion feature vector, the first fault location prediction result and the prediction result probability distribution to obtain the location error of the preset fault location model comprises: performing sequence conversion and alignment operation on the fusion feature vector and the first fault location prediction result by a linear interpolation algorithm to obtain an aligned fusion feature sequence and an aligned location prediction result sequence; and obtaining the location error according to the aligned fusion feature sequence, the aligned location prediction result sequence and the prediction result probability distribution.

[0008] Preferably, the location error is obtained by the following formula:

[0009] is the location error, is a total time of collecting the first data set, is a feature value of the aligned fusion feature sequence at the t th time, is a fault location probability value of the aligned location prediction result sequence at the t th time, is a norm, is a preset coefficient, is a Jensen-Shannon divergence, is a second fault location prediction result of the i th second data set, is an actual fault location result of the i th second data set, is a prediction result probability distribution of the second fault location prediction result and the actual fault location result.

[0010] ​Preferably, the inputting the first data set into the preset fault location model to obtain a first fault location prediction result of the first data set output by the preset fault location model comprises: performing multi-modal feature alignment and splicing on the first data set to obtain a unified feature vector; inputting the unified feature vector into a spatial correlation perception layer, a time feature extraction layer and a space-time interaction attention layer in the preset fault location model in parallel to obtain a spatial correlation feature matrix output by the spatial correlation perception layer, a time feature tensor output by the time feature extraction layer and a space-time fusion weight matrix output by the space-time interaction attention layer; and performing space-time feature reconstruction and collaborative decision on the spatial correlation feature matrix, the time feature tensor and the space-time fusion weight matrix through a multi-source information fusion mechanism to obtain the first fault location prediction result.

[0011] Preferably, the adjusting parameters and weights of the preset fault location model according to the positioning error to obtain an optimized fault location model comprises: determining an adjustment direction of the preset fault location model by monitoring a change relationship between the positioning error and the parameters of the preset fault location model; obtaining a parameter adjustment amount and a weight adjustment amount of the preset fault location model according to the adjustment direction and a preset disturbance amount; and optimizing the parameters and weights of the preset fault location model according to the parameter adjustment amount and the weight adjustment amount to obtain the optimized fault location model.

[0012] Preferably, the optimizing the first fault location prediction result by using the optimized fault location model to obtain an optimized fault location result comprises: performing consistency checking and positioning accuracy correction on the first fault location prediction result by using the optimized fault location model to obtain the optimized fault location result.

[0013] A specific technical scheme of the second embodiment of the present application is: a fault positioning system of a power distribution network, the system comprising: a data acquisition module, a prediction result output module, a fusion feature vector acquisition module, a positioning error acquisition module, a model optimization module and a positioning result optimization module; the data acquisition module is used to collect a first data set of the power distribution network in a current fault state through an edge side; the data set comprises electrical quantity data, time sequence feature data, load characteristic data and equipment state data; the prediction result output module is used to input the first data set into a preset fault positioning model to obtain a first fault positioning prediction result of the first data set output by the preset fault positioning model; the fusion feature vector acquisition module is used to acquire a standardized feature matrix corresponding to the first data set, to perform weighted summation and information fusion analysis on the standardized feature matrix, and to obtain a fusion feature vector corresponding to the first data set; the positioning error acquisition module is used to acquire a prediction result probability distribution of the preset fault positioning model, and to perform bias quantization evaluation according to the fusion feature vector, the first fault positioning prediction result and the prediction result probability distribution to obtain a positioning error of the preset fault positioning model; the model optimization module is used to adjust parameters and weights of the preset fault positioning model according to the positioning error to obtain an optimized fault positioning model; and the positioning result optimization module is used to optimize the first fault positioning prediction result by using the optimized fault positioning model to obtain an optimized fault positioning result.

[0014] A specific technical scheme of the third embodiment of the present application is: a fault positioning device of a power distribution network, comprising a memory and a processor, the memory storing a computer program, and the computer program being executed by the processor to make the processor execute the steps of the method according to any one of the first embodiment of the present application.

[0015] A specific technical scheme of the fourth embodiment of the present application is: a computer readable storage medium storing a computer program, and the computer program being executed by a processor to make the processor execute the steps of the method according to any one of the first embodiment of the present application.

[0016] The embodiments of the present application have the following beneficial effects: The present application generates a standardized feature matrix based on multi-dimensional data such as electrical quantity, time sequence feature, load characteristic and equipment state, and performs weighted fusion and information fusion analysis to obtain a fusion feature vector, so that the fusion feature vector can fully reflect the key criterion information under the fault scene; the positioning error is calculated in combination with the prediction result probability distribution, the fusion feature vector and the first fault positioning prediction result, and the fault positioning model is optimized through the positioning error to form a closed-loop optimization mechanism, so that the fault positioning model has the ability of online learning and self-adaptive evolution, and the robustness and accuracy of the fault positioning model in complex scenes are significantly improved to obtain accurate fault positioning results. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of these drawings.

[0018] Figure 1 The step flow chart of the fault location method for the power distribution network; Figure 2 The structural schematic diagram of the fault location system for the power distribution network; Among them, 201, data acquisition module; 202, prediction result output module; 203, fusion feature vector acquisition module; 204, positioning error acquisition module; 205, model optimization module; 206, positioning result optimization module. DETAILED DESCRIPTION

[0019] The technical solutions in the embodiments of the present application will be described clearly and completely in the following with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments only constitute some embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0020] The terms "first", "second", etc. in the specification and claims and drawings of the present application are used to distinguish different objects, rather than to describe a specific order. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device including a series of steps or modules is not limited to the listed steps or modules, but can optionally include steps or modules not listed, or can optionally include other steps or modules inherent to the process, method, product or device.

[0021] In this paper, the term "embodiment" means that the specific features, structures or characteristics described in conjunction with the embodiment can be included in at least one embodiment of the present application. The phrase appears in the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment independent of or alternative to other embodiments. Those skilled in the art explicitly and implicitly understand that the embodiments described herein can be combined with other embodiments.

[0022] Please refer to Figure 1For the first embodiment of the power distribution network fault location method, a flowchart is provided to obtain accurate fault location results. The method includes: Step 101, collecting a first data set of the power distribution network in the current fault state through the edge side; the data set includes electrical quantity data, time series feature data, load characteristic data, and device state data; Step 102, inputting the first data set into a preset fault location model to obtain a first fault location prediction result of the first data set output by the preset fault location model; Step 103, obtaining a standardized feature matrix corresponding to the first data set, performing weighted summation and information fusion analysis on the standardized feature matrix to obtain a fusion feature vector corresponding to the first data set; Step 104, obtaining a prediction result probability distribution of the preset fault location model, and performing bias quantification evaluation according to the fusion feature vector, the first fault location prediction result, and the prediction result probability distribution to obtain a positioning error of the preset fault location model; Step 105, adjusting the parameters and weights of the preset fault location model according to the positioning error to obtain an optimized fault location model; Step 106, optimizing the first fault location prediction result using the optimized fault location model to obtain an optimized fault location result.

[0023] Specifically, when a fault occurs in a certain power distribution network area, a first data set in the current fault state is collected through an intelligent terminal deployed on the edge side. The collected first data set is input into a preset fault location model, which is constructed based on a deep learning algorithm and trained with a large amount of historical fault data. The model outputs a first fault location prediction result of the first data set, which preliminarily judges the possible location of the fault. The first data set is standardized to obtain a standardized feature matrix, which is weighted and summed. Different types of data are given different weights according to their importance to fault location, and then information fusion analysis is performed to generate a fusion feature vector. The prediction result probability distribution of the preset fault location model is obtained to reflect the confidence of the model in predicting different fault locations. According to the fusion feature vector, the first fault location prediction result, and the prediction result probability distribution, the bias quantification evaluation is performed to calculate the positioning error of the preset fault location model. According to the positioning error, the parameters and weights of the preset fault location model are adjusted using an optimization algorithm to obtain an optimized fault location model. The first fault location prediction result is analyzed and optimized again using the optimized fault location model to obtain the final optimized fault location result, improving the accuracy and reliability of fault location.

[0024] The method in the embodiment generates a standardized feature matrix based on multi-dimensional data such as electrical quantities, timing characteristics, load characteristics, and device states, and performs weighted fusion and information fusion analysis to obtain a fusion feature vector, so that the fusion feature vector can fully reflect the key criterion information in the fault scenario; the positioning error is calculated by combining the prediction result probability distribution, the fusion feature vector, and the first fault location prediction result, and the fault location model is optimized through the positioning error, forming a closed-loop optimization mechanism, so that the fault location model has the ability of online learning and self-adaptive evolution, significantly improving the robustness and accuracy of the fault location model in complex scenarios, to obtain accurate fault location results.

[0025] In specific embodiments, the method comprises: obtaining a plurality of second data sets of the power distribution network in different historical fault states, wherein one second data set corresponds to one historical fault state; inputting the second data set into the preset fault location model to obtain a second fault location prediction result output by the preset fault location model for each second data set; and obtaining the prediction result probability distribution of the preset fault location model according to the second fault location prediction result and the actual fault location result of different historical fault states.

[0026] Specifically, in the power distribution network fault location scenario, in order to obtain the prediction result probability distribution of the preset fault location model, a plurality of second data sets of the power distribution network in different historical fault states are collected. These data sets cover a variety of fault scenarios, such as short-circuit faults and ground faults, and one second data set corresponds to one independent historical fault state. The data content includes electrical quantities, timing characteristics, load characteristics, and device states. The collected second data sets are sequentially input into the preset fault location model. The model is constructed based on advanced algorithms and has fault location capability after a large amount of pre-training. After the model analyzes and processes each second data set, the corresponding second fault location prediction result is output, which clearly indicates the predicted fault location. According to the second fault location prediction result and the actual fault location result of different historical fault states, statistical analysis is performed. By comparing the differences between the prediction result and the actual result, the frequency of occurrence at different prediction positions and other conditions are counted, so as to construct the prediction result probability distribution of the preset fault location model. This distribution can intuitively present the possibility of the model's prediction of different fault positions. By obtaining the prediction result probability distribution of the preset fault location model, the performance characteristics of the model can be more comprehensively and deeply understood. On the one hand, the accuracy and reliability of the model in predicting different fault positions can be clearly known, providing quantitative indicators for evaluating the pros and cons of the model. On the other hand, in the subsequent bias quantification evaluation link, the probability distribution can be combined to more accurately calculate the positioning error, and then the model parameters and weights can be adjusted accordingly, effectively optimizing the fault location model and improving the accuracy and stability of fault location, ensuring the safe and reliable operation of the power distribution network.

[0027] In specific embodiments, the bias quantitative evaluation according to the fusion feature vector, the first fault location prediction result and the prediction result probability distribution obtains a location error of the preset fault location model, which comprises: performing sequence conversion and alignment operation on the fusion feature vector and the first fault location prediction result by a linear interpolation algorithm to obtain an aligned fusion feature sequence and an aligned location prediction result sequence; and obtaining the location error according to the aligned fusion feature sequence, the aligned location prediction result sequence and the prediction result probability distribution. Specifically, sequence conversion and alignment are realized by a linear interpolation algorithm, which can effectively solve the problem of inconsistent dimensions and sampling frequencies between different data, and ensure the accuracy and rationality of subsequent evaluation. The location error is calculated based on the aligned data and the prediction result probability distribution, which can comprehensively consider the accuracy and reliability of model prediction, so that the evaluation result of the location error is closer to the actual situation.

[0028] In specific embodiments, the location error is obtained by the following formula:

[0029] wherein, is the location error, is the total time for collecting the first data set, is a feature value of the aligned fusion feature sequence at the t th time, is a fault location probability value of the aligned location prediction result sequence at the t th time, is a norm, is a preset coefficient, is a Jensen-Shannon divergence, is a second fault location prediction result of the i th second data set, is an actual fault location result of the i th second data set, is a prediction result probability distribution of the second fault location prediction result and the actual fault location result. Specifically, by calculating the difference between the aligned fusion feature sequence and the aligned location prediction result sequence at each time and taking the average, the prediction accuracy of the model for fault features and location results at different times is comprehensively evaluated, avoiding the one-sidedness of single-time evaluation, and the performance of the model in the entire data collection period can be more truly reflected. The Jensen-Shannon divergence can measure the difference between two probability distributions, and by including it in the location error calculation, not only the prediction result itself is concerned, but also the reliability of the prediction result is considered. By comprehensively considering the feature difference and the probability distribution difference to calculate the location error, the location error of the model can be more accurately quantified.

[0030] In specific embodiments, the inputting the first data set into a preset fault location model to obtain a first fault location prediction result of the first data set output by the preset fault location model comprises: performing multi-modal feature alignment and splicing on the first data set to obtain a unified feature vector; inputting the unified feature vector into a spatial correlation perception layer, a time feature extraction layer and a space-time interaction attention layer in the preset fault location model in parallel to obtain a spatial correlation feature matrix output by the spatial correlation perception layer, a time feature tensor output by the time feature extraction layer and a space-time fusion weight matrix output by the space-time interaction attention layer; and performing space-time feature reconstruction and collaborative decision making on the spatial correlation feature matrix, the time feature tensor and the space-time fusion weight matrix through a multi-source information fusion mechanism to obtain the first fault location prediction result.

[0031] Specifically, after obtaining the first data set collected on the edge side, multi-modal feature alignment and splicing are performed thereon. Since the data set contains data of different modalities such as electrical quantity data and time series feature data, a specific alignment algorithm is used to make the modal data consistent in time and space dimensions, and then the aligned features are spliced to form a unified feature vector. The unified feature vector is input into different functional layers of the preset fault location model in parallel. Among them, the spatial correlation perception layer focuses on analyzing the spatial correlation between distribution network devices, outputs a spatial correlation feature matrix, the time feature extraction layer focuses on mining the change law of data over time, outputs a time feature tensor, and the space-time interaction attention layer focuses on the interaction between spatial and temporal features, and outputs a space-time fusion weight matrix. A multi-source information fusion mechanism is used to comprehensively process the spatial correlation feature matrix, the time feature tensor and the space-time fusion weight matrix. This mechanism fully considers the correlation and complementarity between features, reconstructs space-time features, and comprehensively judges the possible location of the fault through collaborative decision making, and finally outputs the first fault location prediction result.

[0032] In specific embodiments, the adjusting parameters and weights of the preset fault location model according to the positioning error to obtain an optimized fault location model comprises: determining an adjustment direction of the preset fault location model by monitoring the change relationship between the positioning error and the parameters of the preset fault location model; obtaining a parameter adjustment amount and a weight adjustment amount of the preset fault location model according to the adjustment direction and a preset disturbance amount; and optimizing the parameters and weights of the preset fault location model according to the parameter adjustment amount and the weight adjustment amount to obtain the optimized fault location model.

[0033] Specifically, in the optimization process of the power distribution network fault location model, when the positioning error of the preset fault location model is obtained, the parameters and weights thereof are adjusted. The adjustment direction is determined by monitoring the change relationship between the positioning error and the model parameters in real time. After the adjustment direction is determined, the parameter adjustment amount and the weight adjustment amount are calculated in combination with the preset disturbance amount. The preset disturbance amount is a small change amount set according to the model characteristics and actual requirements, and the specific adjustment value is obtained according to the adjustment direction and the disturbance amount. The parameters and weights of the preset fault location model are updated and optimized according to the calculated parameter adjustment amount and weight adjustment amount. After a series of adjustments, the optimized fault location model is obtained, which is more suitable for actual requirements in terms of parameter and weight configuration, and can more accurately locate the fault of the power distribution network.

[0034] In specific embodiments, the optimization of the first fault location prediction result by using the optimized fault location model to obtain an optimized fault location result comprises consistency checking and positioning accuracy correction of the first fault location prediction result by using the optimized fault location model to obtain an optimized fault location result. Specifically, consistency checking can find unreasonable places in the first fault location prediction result and ensure the logical rationality of the result. Positioning accuracy correction directly improves the accuracy of fault location. Through these two steps, the advantages of the optimized fault location model are fully utilized to optimize the preliminary prediction result, effectively reduce the error of fault location, and improve the reliability and accuracy of power distribution network fault location, which helps power operation and maintenance personnel to handle faults faster and more accurately and ensures the stable operation of the power system.

[0035] Specifically, the detailed embodiments of the method are as follows: S1, collecting multi-dimensional data in the power distribution network in real time through the edge side, and preprocessing the multi-dimensional data.

[0036] The multi-dimensional data includes electrical quantity data, time sequence characteristic data, load characteristic data and equipment state data. It should be noted that the electrical quantity data refers to the basic physical parameters collected by sensors in real time, which are used to directly represent the power quality and energy flow status of the power distribution network, mainly including real-time measured voltage, current, power, frequency and phase basic power parameters. The time sequence characteristic data refers to the characteristic data with time sequence and dynamic change law extracted from the electrical quantity data and the measured values with time sequence attributes and reflecting the state change of the power grid, which is used to analyze the type, occurrence and development process of the fault. The load characteristic data refers to the data related to the power consumption behavior and time distribution of the load in the power distribution network, which is used to distinguish the real power grid fault from the normal disturbance caused by the switching of large-capacity load. The equipment state data refers to the information reflecting the operating state of circuit breakers, disconnectors and relay protection electrical equipment in the power distribution network, which is used to determine the specific section and impact range of the fault.

[0037] The preprocessing includes denoising filtering, time alignment, normalization and feature extraction. It should be pointed out that the denoising filtering refers to using a digital filter to perform signal smoothing and noise removal operation on the electrical quantity data and time sequence characteristic data collected in real time on the edge side, remove measurement white noise and transient pulse interference, and retain the true electrical change trend; the time alignment refers to the operation of unifying the data from different collection terminals and the multi-source data with millisecond timestamp difference to the same standard time axis through interpolation and resampling, ensuring the comparability of the data at the same time; the normalization refers to the operation of mapping the electrical quantity data and load characteristic data to the dimensionless interval [0, 1] and [-1, 1] through linear and nonlinear transformation, so as to remove the dimension influence and provide a stable foundation for the training of the fault location model; the feature extraction refers to the operation of the multi-dimensional data after denoising filtering, time alignment and normalization through statistical methods, signal transformation and model algorithms, to generate a low-dimensional feature vector that can essentially represent the power grid operation state and fault characteristics.

[0038] S2, constructing a fault location model, inputting the multi-dimensional data after preprocessing into the fault location model, performing power distribution network fault identification and preliminary fault location, and generating a preliminary fault location result.

[0039] Based on the graph attention network, a spatial correlation perception layer is constructed, based on the one-dimensional time sequence convolution network, a time feature extraction layer is constructed, and based on the attention mechanism, a space-time interaction attention layer is constructed.

[0040] Specifically, the power distribution network topology graph is abstracted according to the physical connection relationship between electrical equipment in the power distribution network, the node represents the power distribution equipment, and the edge represents the electrical connection; the electrical quantity data and equipment state data of the nodes in the neighborhood of each node are aggregated through the graph attention network to generate a node feature representation containing spatial dependency, the node feature representations are collected into a spatial correlation feature matrix, and the spatial correlation perception layer is constructed.

[0041] The time sequence of the aligned electrical quantity data and load characteristic data is scanned through the one-dimensional time sequence convolution network to extract local trend and transient mode features at different time scales, the local trend and transient mode features are weighted and fused to form a unified time feature tensor, and the construction of the time feature extraction layer is completed.

[0042] The node features of the spatial correlation perception layer and the time feature tensor of the time feature extraction layer are cross-dimensionally associated and mapped to obtain a space-time dimension distribution, the feature combination in the space-time dimension distribution that is strongly related to the fault is selected through feature interaction, a space-time fusion weight matrix is generated to complete the construction of the space-time interaction attention layer.

[0043] It should be noted that the pre-training process of the graph attention network is as follows: a power distribution network simulation environment is constructed on a simulation platform, historical fault data is used as input, the accuracy of fault location is used as a supervision signal, and the parameters of the graph attention network are optimized through a back propagation algorithm, so that the graph attention network can learn the spatial correlation features between nodes in the power distribution network. The pre-training process of the one-dimensional time sequence convolution network is as follows: a power distribution network simulation environment is constructed on a simulation platform, time series in historical fault data are used as input, the accuracy of fault location is used as a supervision signal, and the convolution kernel weight and bias parameters of the one-dimensional time sequence convolution network are optimized through a back propagation algorithm, so that the one-dimensional time sequence convolution network can learn the effective time features in the power data.

[0044] The spatial correlation perception layer, the time feature extraction layer, and the space-time interaction attention layer are spliced to form a fault location model.

[0045] Specifically, the spatial correlation feature matrix generated by the spatial correlation perception layer, the time feature tensor generated by the time feature extraction layer, and the space-time fusion weight matrix generated by the space-time interaction attention layer are cascaded and weighted aggregated through a multi-source information fusion mechanism to form a unified deep feature representation. The deep feature representation is mapped through a full connection network to obtain a probability distribution of the fault area, and a fault location model is formed.

[0046] It should be noted that the pre-training process of the fault location model is as follows: a power distribution network digital simulation environment is constructed on a simulation platform, fault scenarios of different fault types, fault locations, and operating modes are simulated, training samples containing multi-dimensional electrical quantity data and corresponding fault location labels are generated, the training samples are input into the fault location model for learning, the error is obtained by comparing the fault location output by the fault location model with the accurate fault location label in the simulation platform, and the internal parameters of the fault location model are repeatedly adjusted according to the error until the fault location model can stably reach the predetermined positioning accuracy requirement in the test of new fault scenarios not participating in the training, thereby completing the pre-training and obtaining the pre-trained fault location model.

[0047] The predetermined positioning accuracy is calculated by K-fold cross-validation of historical fault data set in the fault positioning model pre-training stage, and the performance threshold is set according to the industry reliability standard, which is used to objectively measure the correctness of the model in judging the fault position. The exemplary value range is 95% to 99%. The value range is based on the general reliability requirement of the power system protection industry for the correct action rate of key fault detection equipment (not less than 95%), and combined with the historical data verification that the model performance reaches the critical point of diminishing returns (more than 99% after which the cost is too high and may be over-fitted) to determine the industry consensus range. When the value is less than 95%, the fault positioning model does not meet the reliability requirement, and the fault positioning model parameters need to be adjusted until the threshold requirement is met. When the value is higher than 99%, the fault positioning model has reached the optimal performance and can be accepted.

[0048] The preprocessed multi-dimensional data is subjected to multi-modal feature alignment and splicing to form a unified feature vector.

[0049] Specifically, the time alignment operation is performed on the preprocessed electrical quantity data, preprocessed time sequence feature data, preprocessed load characteristic data and preprocessed equipment state data to generate aligned multi-modal data with a unified timestamp sequence. The aligned multi-modal data is converted into a preset fixed-dimension electrical quantity sub-vector, time sequence feature sub-vector, load characteristic sub-vector and equipment state sub-vector through feature mapping operation, and vector splicing operation is performed to output a unified feature vector.

[0050] It should be noted that the preset fixed dimension refers to the pre-defined length of each feature sub-vector, which is determined according to the specific output dimension of each data type in the fault positioning model architecture, to ensure that all sub-vectors can be subjected to vector splicing operation.

[0051] The unified feature vector is input into the spatial correlation perception layer, the time feature extraction layer and the space-time interaction attention layer in parallel to output the spatial correlation feature matrix, the time feature tensor and the space-time fusion weight matrix respectively.

[0052] Specifically, the spatial correlation perception layer aggregates the adjacent node features of each node information by fusing the distribution network topology graph and the node information in the unified feature vector to generate a spatial correlation feature matrix containing global spatial dependency; the time feature extraction layer reorganizes the unified feature vector into time sequence data in time sequence, and then applies a one-dimensional time sequence convolution network to perform deep convolution and pooling operation on the time sequence data to obtain a time feature tensor; the space-time interaction attention layer performs tensor multiplication and Softmax transformation on the spatial correlation feature matrix and the time feature tensor to generate an attention distribution, and performs weighted summation on the feature vectors at different time steps in the time feature tensor to generate a space-time fusion weight matrix.

[0053] The spatial correlation feature matrix, the time feature tensor and the spatiotemporal fusion weight matrix are reconstructed and cooperatively decided through a multi-source information fusion mechanism to generate a preliminary fault location result.

[0054] Specifically, the spatial correlation feature matrix and the time feature tensor are spliced to form a spatiotemporal joint feature, the spatiotemporal fusion weight matrix is used as a weight coefficient to weight and fuse the spatiotemporal joint feature, and a reconstructed deep spatiotemporal feature is generated; the full connection layer maps the deep spatiotemporal feature to a non-normalized logarithmic probability, which is then converted to a fault location probability distribution through a Softmax function, and the equipment node with the highest fault location probability value in the fault location probability distribution is determined as the fault point to generate a preliminary fault location result.

[0055] It should be noted that the equipment node refers to a basic object defined for abstractly describing the topology structure of the power distribution network when constructing the fault location model, which is created by mapping the equipment identifier to a graph node in the fault location model initialization stage according to the actual monitored physical equipment in the power distribution network.

[0056] S3, receiving the preprocessed multi-dimensional data through the cloud platform, and performing time alignment and normalization to obtain a standardized feature matrix, and performing weighted summation and information fusion analysis on the standardized feature matrix through the weighted average method to obtain a fusion feature vector.

[0057] The timestamp field of the preprocessed multi-dimensional data is extracted for sorting and removing duplicates to form a unique time point sequence, and the minimum common time interval is determined according to the interval difference between adjacent time points in the unique time point sequence to generate a uniform time grid.

[0058] Specifically, the timestamp field of the preprocessed multi-dimensional data is extracted and arranged in ascending order to generate an ordered timestamp sequence, and the duplicates of the ordered timestamp sequence are removed through a hash deduplication operation to generate a unique time point sequence, the interval difference between adjacent time points in the unique time point sequence is obtained and the minimum interval difference is selected to obtain the minimum common time interval, and the first time point of the unique time point sequence is taken as the starting point and the minimum common time interval is taken as the fixed step size to perform an arithmetic sequence generation operation to generate a uniform time grid.

[0059] The preprocessed multi-dimensional data is aligned to the uniform time grid through a linear interpolation method to form a time-aligned data matrix.

[0060] Specifically, according to the target time point in the uniform time grid, two adjacent data points on the time sequence of the preprocessed multi-dimensional data are found, and linear interpolation calculation is performed on the two data points and the corresponding timestamp positions to obtain the interpolation data of the target time point; the interpolation data of all target time points are arranged in chronological order to form a time-aligned data matrix.

[0061] It should be noted that the target time point is each equally spaced time point predefined and determined by the unified time grid, which is used as a reference to align the pretreated multi-dimensional data to the unified time axis by linear interpolation. The data matrix after time alignment is adjusted by center translation and scaling to obtain the standardized feature matrix.

[0062] Specifically, the center translation adjustment operation is performed on the data matrix after time alignment to obtain the arithmetic mean of each column of data points in the data matrix, and the centralized data matrix is formed according to the difference between each column of data points and the corresponding arithmetic mean. The scaling adjustment operation is performed on the centralized data matrix to obtain the standard deviation of each column of features, and each column of data points is divided by the corresponding standard deviation as the denominator to obtain the standardized feature matrix. The deviation matrix is obtained by analyzing the deviation degree of the standardized feature matrix.

[0063] Specifically, the absolute difference between each data point in the standardized feature matrix and the preset reference value is obtained by element-by-element subtraction and absolute value operation, and the absolute difference is filled into a matrix with the same dimension as the standardized feature matrix to form the deviation matrix.

[0064] It should be noted that the preset reference value is set based on the mathematical properties of the standardized feature matrix, and the exemplary value is 0, which is determined according to the data center position of the standardized feature matrix. Comparing each data point with 0 can intuitively measure the degree of deviation from the data center.

[0065] The deviation matrix is enhanced and integrated by exponential enhancement operation to generate unnormalized adjustment weight, and the normalized weight matrix is obtained by normalizing the unnormalized adjustment weight. Specifically, each element in the deviation matrix is nonlinearly amplified by an exponential function with a natural constant as the base to generate unnormalized adjustment weight. Each unnormalized adjustment weight is taken as the numerator, and the sum of all elements of the unnormalized adjustment weight is taken as the denominator to perform division operation, so that each element in the deviation matrix is converted into the proportion of the sum of all elements in the unnormalized adjustment weight, and the normalized weight matrix with the sum of all elements being 1 is obtained.

[0066] It should be noted that by using the monotone increasing and non-negative characteristics of the exponential function, the deviation value representing the deviation degree is nonlinearly mapped to a positive weight, and the greater the deviation degree, the greater the amplification of the weight.

[0067] The weighted sum is performed on the normalized weight matrix and the standardized feature matrix to generate a fusion feature value, and the fusion feature values of each time point are arranged and combined in time sequence to generate a fusion feature vector. Specifically, according to the time points of the unified time grid, the weight row vector of the corresponding time point in the normalized weight matrix and the feature row vector of the corresponding time point in the standardized feature matrix are extracted, and the weight row vector and the feature row vector are element-wise weighted and summed to generate a fusion feature value of the time point. The fusion feature values of all time points are arranged in time sequence to form a fusion feature vector.

[0068] S4, bias quantitative evaluation is performed on the fusion feature vector and the preliminary fault location result according to the dynamic error evaluation algorithm to obtain a positioning error.

[0069] The fusion feature vector and the preliminary fault location result are subjected to sequence conversion and alignment operation by a linear interpolation algorithm to obtain an aligned fusion feature sequence and an aligned positioning result sequence. Specifically, the timestamp sequence of the fusion feature vector is selected as a reference time axis, and for each target time point on the reference time axis, the fault location probability values of two adjacent data points of the target time point are found on the time sequence of the preliminary fault location result, and the fault location probability values of the two data points and the corresponding time positions are used for linear interpolation calculation to obtain the fault location probability interpolation value of the target time point. The fault location probability interpolation values of all target time points are arranged in time sequence to form an aligned positioning result sequence completely aligned with the reference time axis. Since the timestamp sequence of the fusion feature vector is consistent with the reference time axis, it is directly used as the aligned fusion feature sequence.

[0070] It should be noted that the fault location probability value refers to a value between 0 and 1 calculated by the fault location model for each device node in the distribution network, which is derived from the preliminary fault location result generated by the fault location model after processing the unified feature vector.

[0071] The aligned fusion feature sequence and the aligned positioning result sequence are compared point by point to generate an absolute deviation sequence.

[0072] Specifically, the aligned fusion feature sequence and the aligned positioning result sequence have the same timestamp sequence, and according to each time point in the timestamp sequence, the fault location probability value corresponding to each time point in the aligned positioning result sequence and the feature value corresponding to each time point in the aligned fusion feature sequence are extracted to obtain the absolute difference between the positioning probability value and the feature value of each time point. The absolute difference values of all time points in the timestamp sequence are arranged in time sequence to generate an absolute deviation sequence, and the deviation sequence is combined with the prediction probability result to obtain the positioning error, i.e., the positioning error is obtained using the following formula:

[0073] wherein, is the positioning error, is the total time for collecting the first data set, is the feature value of the aligned fusion feature sequence at the t-th time, is the fault location probability value of the aligned positioning prediction result sequence at the t-th time, is a norm, is a preset coefficient, is a Jensen-Shannon divergence, is the second fault location prediction result of the i-th second data set, is the actual fault location result of the i-th second data set, is the prediction result probability distribution of the second fault location prediction result and the actual fault location result.

[0074] S5, adjusting the parameters and weights of the fault location model according to the positioning error, optimizing the preliminary fault location result to generate an optimized fault location result, and performing consistency checking and impact range determination on the optimized fault location result and the operation parameters of the distribution network to obtain a distribution network fault location report.

[0075] By monitoring the change relationship between the positioning error and the fault location model parameters, the adjustment direction of the fault location model parameters is determined. Specifically, according to the positioning error, each parameter in the fault location model is adjusted in turn by a preset small perturbation, and the change amount of the positioning error after each parameter adjustment is observed; by comparing the size and direction of the change amount of the positioning error caused by the small adjustment of each parameter, the adjustment direction of the fault location model parameters is determined, when the increase of the fault location model parameters leads to the decrease of the positioning error, the adjustment direction of the fault location model parameters is positive increase; when the increase of the fault location model parameters leads to the increase of the positioning error, the adjustment direction of the fault location model parameters is negative decrease.

[0076] It should be noted that the preset small perturbation is set according to the numerical stability requirement of gradient estimation, and an exemplary value is 0.001, which is determined according to the numerical differentiation principle and the scale value of the model parameters.

[0077] According to the adjustment direction of the fault location model parameters, the fault location model parameters and weights are adaptively adjusted to obtain an adjustment amount.

[0078] Specifically, based on the adjustment direction of the fault location model parameters, preset small disturbance quantities are assigned positive or negative signs, with the signs consistent with the adjustment direction (e.g., a positive sign is assigned when the adjustment direction is positively increasing, and a negative sign is assigned when the adjustment direction is negatively decreasing). The signed small disturbance quantities are directly used as the adjustment amounts of the fault location model parameters. The weights are part of the fault location model parameters, and the adjustment amounts of the weights are consistent with the adjustment amounts of the corresponding fault location model parameters, and adaptive adjustment operations are performed synchronously.

[0079] The fault location model is updated based on the adjustment amount. The updated fault location model is then used to verify the consistency of the preliminary fault location results and correct the location accuracy, thereby generating optimized fault location results.

[0080] Specifically, the current value of each parameter in the fault location model is added to its corresponding adjustment amount, and the result directly replaces the current value of each parameter in the fault location model to obtain the updated fault location model. The preliminary fault location result is then re-inputted into the updated fault location model to obtain a new fault location probability distribution. The new fault location probability distribution is then compared point by point with the fault location probability distribution of the preliminary fault location result to obtain the probability deviation of each device node. When the probability deviation of a device node exceeds a preset probability deviation threshold, the probability value in the preliminary fault location result is directly replaced with the new probability value output by the updated fault location model to generate an optimized fault location result.

[0081] It should be noted that the preset probability deviation threshold is determined by statistically analyzing the probability difference distribution between the preliminary and optimized location results in a large number of correct and incorrect location cases. This threshold is a critical value that can effectively distinguish between "normal fluctuations" and "significant deviations." An exemplary value range is 0.1 to 0.3. The value is determined by analyzing the statistical characteristics of the probability difference distribution in a large number of cases (such as calculating the mean and standard deviation, and determining the optimal critical point based on the ROC curve) to ensure that normal fluctuations and significant deviations are effectively distinguished while minimizing the risk of false alarms and false negatives. When the probability deviation is less than 0.1, the probability change belongs to the normal fluctuation after the fault location model is optimized, and no correction is made. The original probability value in the preliminary fault location result is retained. When the probability deviation is greater than 0.3, the preliminary location result has a deviation and is corrected. The new probability value output by the updated fault location model replaces the original probability value.

[0082] Based on the optimized fault location results, the equipment in the distribution network is mapped as graph nodes, and electrical connections are mapped as edges, forming a graph structure. Specifically, based on the optimized fault location results, all monitored equipment object identifiers are extracted from the preset distribution network monitoring database. Each equipment object identifier is instantiated as a node in the graph data structure. The fault probability value corresponding to each device is retrieved by matching the equipment identifier, and the fault probability value is assigned as the "fault probability" attribute of the corresponding graph node. The electrical connection relationships recorded in the preset distribution network topology connection relationship database are converted into undirected edges established between the corresponding nodes in the graph data structure, forming a graph structure containing node attributes and connection relationships.

[0083] It should be noted that the preset distribution network monitoring database refers to the database that has been established in advance before the implementation of the fault location method, including equipment object identifiers and equipment status data; the preset distribution network topology connection relationship database refers to the database that has been established in advance before the implementation of the fault location method, including electrical connection relationships and network topology diagrams.

[0084] The optimized fault location result is compared with the preset expected fault location result to obtain the location deviation. When the location deviation exceeds the preset consistency threshold, a consistency evaluation result is generated.

[0085] Specifically, for each device node, the absolute difference between the optimized fault location result and the preset expected fault location result is obtained to obtain the location deviation of the device node. The location deviation of each device node is compared with the preset consistency threshold: when the location deviation of the device node is higher than the preset consistency threshold, the device node is marked as "inconsistent" and when the location deviation of the device node is lower than the preset consistency threshold, it is marked as "consistent". The set of the marked states of all device nodes constitutes the consistency evaluation result.

[0086] It should be noted that the preset consistency threshold is set by statistically analyzing the fluctuation range of characteristic parameters in historical normal operation data and setting the consistency threshold to a specific percentile based on the acceptable false alarm rate for the business. An exemplary value range is 0.1 to 0.3. A value below 0.1 would make the consistency condition too strict, easily misjudging normal fluctuations as abnormal fault location, resulting in low report fault tolerance; a value above 0.3 would make the consistency condition too lenient, omitting true location deviations and reducing report reliability.

[0087] Topological connectivity analysis and impact diffusion simulation were performed on the graph structure and consistency assessment results to obtain the impact range assessment results.

[0088] Specifically, the device nodes marked as "inconsistent" in the consistency assessment results are taken as the starting point. All adjacent nodes on the graph structure are visited using a breadth-first search algorithm. All visited nodes are organized into a node set, and the node set is directly defined as the scope of the device affected by the fault deviation at the starting point as the scope of the impact assessment result.

[0089] The consistency assessment results and impact range assessment results are structurally integrated and timestamped to generate a distribution network fault location report. Specifically, the marked status of equipment nodes in the consistency assessment results are correlated and matched with the impact range assessment results to form equipment-level assessment items. All equipment-level assessment items are arranged according to a preset report format to form the main body of the distribution network fault location report, and the current timestamp at the time of report generation is embedded in the report title, together forming a structured distribution network fault location report.

[0090] It should be noted that the preset report format is a predefined structured data arrangement specification before the report is generated. It is a pre-defined data layout template to meet the business analysis and standardization needs of fault location, ensuring that each generated report has a unified structure, complete information, and highlights key points.

[0091] S6. Based on the distribution network fault location report, the fault location model is iteratively optimized through cloud-edge collaboration.

[0092] The simulation environment is reconstructed on a pre-built biomimetic platform. Based on the power distribution network fault location report, the fault location model is simulated in the simulation environment to generate location simulation results and performance evaluation reports.

[0093] Specifically, based on the power distribution network fault location report, electrical equipment is mapped as simulation nodes and electrical connections are mapped as simulation links. The simulation environment is reconstructed on a pre-set biomimetic platform. After the fault location model is loaded into the simulation environment, preset simulation test data is input and the fault location model is run. The fault location model performs spatiotemporal feature reasoning and probability classification calculation on the fault scenarios in the simulation environment to obtain the fault probability distribution of each equipment node and form the location simulation result.

[0094] The simulation results are compared with the preset fault location results node by node. Based on the comparison results, the number of correctly located, missed, and false alarm nodes is counted. The ratio of the number of correctly located nodes to the total number of nodes is compared with the preset performance compliance threshold. When the ratio is greater than or equal to the preset performance compliance threshold, a qualitative conclusion of "model performance meets the standard" is generated. The proportional relationship between the number of correctly located, missed, and false alarm nodes and the total number of nodes, the actual number of faulty nodes and the number of normal nodes in the preset fault location results is established respectively. Indicators including positioning accuracy, missed alarm rate and false alarm rate are directly generated, and a performance evaluation report containing qualitative conclusions and indicators is generated.

[0095] It should be noted that the preset simulation test data refers to the set of simulation data pre-built on the biomimetic platform for testing the performance of the fault location model, which includes electrical quantity data of each monitoring point in the simulation environment under various fault conditions. Preset fault location results refer to the unique and correct fault location answer predefined for each set of preset simulation test data before the simulation test begins. It serves as a benchmark for evaluating the performance of the fault location model. The preset expected fault location results include the faulty device identifier, fault type, and expected fault probability distribution.

[0096] The preset performance threshold refers to the minimum accuracy standard used to judge whether the performance of the fault location model is qualified. An example value is 95%, which is determined according to the minimum requirements for fault location accuracy for safe operation of the power grid.

[0097] The preset indicator fields refer to the names of data entries that are predefined in the performance evaluation report template and used to quantify the performance of the fault location model, including location accuracy, false alarm rate, false alarm rate, and average location time.

[0098] The cloud platform performs deviation analysis on the location simulation results based on the performance evaluation report and generates parameter adjustment plans. Specifically, the cloud platform cross-compares the performance evaluation report with preset fault type and topology labels to locate the specific fault type and topology scenario causing the performance deviation. By analyzing samples with erroneous location simulation results in the topology scenario, the platform observes the transmission and transformation process of the spatiotemporal characteristics of the multi-dimensional data corresponding to the erroneous samples in each layer of the fault location model, identifies the underperforming components, and determines the fault location parameters that need to be adjusted and the adjustment direction based on the underperforming components and the gradient descent principle, thus generating a parameter adjustment plan.

[0099] It should be noted that the preset fault types and topology labels refer to the standardized classification identifiers predefined for the simulation test scenario. The preset fault types include electrical nature faults, transition resistance faults, and development process faults. The preset topology labels refer to the predefined standardized grid connection relationship identifiers.

[0100] The fault location model on the edge side is updated according to the parameter adjustment scheme, and multi-scenario verification tests are carried out in the same simulation environment until the convergence condition is reached, thus completing the iterative optimization of the fault location model.

[0101] Specifically, the adjustment direction in the parameter adjustment scheme is applied to the corresponding fault location model parameters of the edge-side fault location model to complete the update of the fault location model; the updated fault location model is loaded into the original simulation environment and run multiple test scenarios covering different fault types and topologies. After each test, the performance evaluation report is checked to see if the indicators meet the preset convergence conditions. The test is repeated until the indicators are stable above the preset performance threshold for several consecutive tests, thus completing the iterative optimization of the fault location model.

[0102] It should be noted that the preset convergence condition refers to the quantitative standard for determining that the optimization process of the fault location model can be terminated, requiring the performance indicators of the fault location model to stably reach or exceed the preset performance threshold in multiple consecutive tests.

[0103] In a specific embodiment, please refer to Figure 2 This is a schematic diagram of a fault location system for a power distribution network according to a second embodiment of this application. The system includes: a data acquisition module 201, a prediction result output module 202, a fusion feature vector acquisition module 203, a location error acquisition module 204, a model optimization module 205, and a location result optimization module 206. The data acquisition module 201 is used to collect a first dataset of the power distribution network in the current fault state through the edge side. The dataset includes electrical quantity data, time series feature data, load characteristic data, and equipment status data. The prediction result output module 202 is used to input the first dataset into a preset fault location model to obtain a first fault location prediction result of the first dataset output by the preset fault location model. The fusion feature vector acquisition module 203 is used to acquire... The standardized feature matrix corresponding to the first dataset is weighted and summed, and information fusion analysis is performed on the standardized feature matrix to obtain the fused feature vector corresponding to the first dataset. The positioning error acquisition module 204 is used to acquire the probability distribution of the prediction result of the preset fault positioning model, and to perform deviation quantification evaluation based on the fused feature vector, the first fault positioning prediction result, and the probability distribution of the prediction result to obtain the positioning error of the preset fault positioning model. The model optimization module 205 is used to adjust the parameters and weights of the preset fault positioning model according to the positioning error to obtain an optimized fault positioning model. The positioning result optimization module 206 is used to optimize the first fault positioning prediction result using the optimized fault positioning model to obtain an optimized fault positioning result.

[0104] In this embodiment, the system generates a standardized feature matrix based on multi-dimensional data such as electrical quantities, time-series characteristics, load characteristics, and equipment status. Weighted fusion and information fusion analysis are then performed to obtain a fused feature vector, which fully reflects the key judgment information under fault scenarios. The system calculates the location error by combining the probability distribution of the prediction results, the fused feature vector, and the first fault location prediction result. The fault location model is then optimized using the location error, forming a closed-loop optimization mechanism. This enables the fault location model to have online learning and adaptive evolution capabilities, significantly improving the robustness and accuracy of the fault location model in complex scenarios, thereby obtaining accurate fault location results.

[0105] In a specific embodiment, the third embodiment of this application provides a fault location device for a power distribution network, including a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method as described in any one of the first embodiments of this application.

[0106] In a specific embodiment, the fourth embodiment of this application provides a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method as described in any one of the first embodiments of this application.

[0107] The above embodiments merely illustrate several implementation methods of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

[0108] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments for application in other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A fault location method for a power distribution network, characterized in that, The method includes: The first dataset of the distribution network under the current fault state is collected by the edge side; the dataset includes electrical quantity data, time series characteristic data, load characteristic data and equipment status data; The first dataset is input into a preset fault location model to obtain the first fault location prediction result of the first dataset output by the preset fault location model. Obtain the standardized feature matrix corresponding to the first dataset, perform weighted summation and information fusion analysis on the standardized feature matrix, and obtain the fused feature vector corresponding to the first dataset; Obtain the probability distribution of the prediction results of the preset fault location model, and perform deviation quantification evaluation based on the fused feature vector, the first fault location prediction result and the probability distribution of the prediction results to obtain the location error of the preset fault location model; The parameters and weights of the preset fault location model are adjusted according to the location error to obtain an optimized fault location model; The first fault location prediction result is optimized using the optimized fault location model to obtain the optimized fault location result.

2. The fault location method for a distribution network as described in claim 1, characterized in that, The step of obtaining the probability distribution of the prediction results of the preset fault location model includes: Obtain multiple second datasets of the distribution network under different historical fault states, where each second dataset corresponds to one historical fault state; The second dataset is input into a preset fault location model to obtain the second fault location prediction result for each second dataset output by the preset fault location model. The probability distribution of the prediction results of the preset fault location model is obtained based on the second fault location prediction result and the actual fault location results of different historical fault states.

3. The fault location method for a distribution network as described in claim 2, characterized in that, The step of performing deviation quantification evaluation based on the fused feature vector, the first fault location prediction result, and the probability distribution of the prediction result to obtain the location error of the preset fault location model includes: The fused feature vector and the first fault location prediction result are subjected to sequence transformation and alignment operations using a linear interpolation algorithm to obtain an aligned fused feature sequence and an aligned location prediction result sequence. The positioning error is obtained based on the aligned fused feature sequence, the aligned positioning prediction result sequence, and the probability distribution of the prediction results.

4. The fault location method for a distribution network as described in claim 3, characterized in that, The positioning error is obtained using the following formula: in, The positioning error is... The total time for collecting the first dataset, Let be the feature value of the aligned fused feature sequence at time t. Let be the fault location probability value of the aligned location prediction result sequence at time t. For norm, For preset coefficients, For Jensen-Shannon divergence, This is the second fault location prediction result for the i-th second dataset. This represents the actual fault location result for the i-th second dataset. The probability distribution of the predicted results of the second fault location prediction and the actual fault location results.

5. The fault location method for a distribution network as described in claim 1, characterized in that, The step of inputting the first dataset into a preset fault location model to obtain a first fault location prediction result of the first dataset output by the preset fault location model includes: The first dataset is subjected to multimodal feature alignment and concatenation to obtain a unified feature vector; The unified feature vector is input in parallel into the spatial association perception layer, the temporal feature extraction layer, and the spatiotemporal interaction attention layer in the preset fault location model to obtain the spatial association feature matrix output by the spatial association perception layer, the temporal feature tensor output by the temporal feature extraction layer, and the spatiotemporal fusion weight matrix output by the spatiotemporal interaction attention layer. The first fault location prediction result is obtained by reconstructing and coordinating the spatial correlation feature matrix, the temporal feature tensor, and the spatiotemporal fusion weight matrix through a multi-source information fusion mechanism.

6. The fault location method for a distribution network as described in claim 1, characterized in that, The step of adjusting the parameters and weights of the preset fault location model based on the location error to obtain an optimized fault location model includes: By monitoring the relationship between the positioning error and the parameters of the preset fault location model, the adjustment direction of the preset fault location model can be determined. The parameter adjustment amount and weight adjustment amount of the preset fault location model are obtained according to the adjustment direction and the preset disturbance amount; The parameters and weights of the preset fault location model are optimized based on the parameter adjustment amount and the weight adjustment amount to obtain the optimized fault location model.

7. The fault location method for a distribution network as described in claim 1, characterized in that, The step of optimizing the first fault location prediction result using the optimized fault location model to obtain the optimized fault location result includes: using the optimized fault location model to perform consistency verification and location accuracy correction on the first fault location prediction result to obtain the optimized fault location result.

8. A fault location system for a power distribution network, characterized in that, The system includes: a data acquisition module, a prediction result output module, a fusion feature vector acquisition module, a positioning error acquisition module, a model optimization module, and a positioning result optimization module; The data acquisition module is used to collect the first dataset of the distribution network in the current fault state through the edge side; the dataset includes electrical quantity data, time series characteristic data, load characteristic data and equipment status data; The prediction result output module is used to input the first dataset into the preset fault location model to obtain the first fault location prediction result of the first dataset output by the preset fault location model. The fusion feature vector acquisition module is used to acquire the standardized feature matrix corresponding to the first dataset, perform weighted summation and information fusion analysis on the standardized feature matrix, and obtain the fusion feature vector corresponding to the first dataset. The positioning error acquisition module is used to acquire the probability distribution of the prediction results of the preset fault positioning model, and to perform deviation quantification evaluation based on the fused feature vector, the first fault positioning prediction result and the probability distribution of the prediction results to obtain the positioning error of the preset fault positioning model. The model optimization module is used to adjust the parameters and weights of the preset fault location model according to the location error, so as to obtain an optimized fault location model. The location result optimization module is used to optimize the first fault location prediction result using the optimized fault location model to obtain the optimized fault location result.

9. A fault location device for a power distribution network, comprising a memory and a processor, characterized in that, The memory stores a computer program that, when executed by the processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it causes the processor to perform the steps of the method as described in any one of claims 1 to 7.

Images