Fault identification methods and related devices for power distribution networks

Abstract

This application discloses a fault identification method and related apparatus for a power distribution network. The method includes: acquiring a first output result of a reference spatiotemporal convolutional model; adjusting the feature output dimension of the reference spatiotemporal convolutional model according to the target feature dimension to obtain a target spatiotemporal convolutional model; inputting dynamic data into the target spatiotemporal convolutional model to output target spatiotemporal features, wherein the dynamic data includes the electrical operating status of the power distribution network and the external environmental status; fusing the target spatiotemporal features and static data to obtain initial fault features, wherein the static data includes the topology, electrical parameters, and operation and maintenance records of the power distribution network; determining the first contribution of each feature in the initial fault features; filtering the initial fault features according to the first contribution to obtain target fault features; and performing fault identification based on the target fault features to obtain a fault result. This application can effectively improve the fault identification accuracy while reducing the demand for computing power.

Description

Technical Field

[0001] This application relates to the field of power distribution network fault identification technology, and in particular to a fault identification method and related apparatus for power distribution networks. Background Technology

[0002] As the core terminal link of the power system, the distribution network plays a crucial role in distributing electrical energy from the transmission network to industrial, commercial, and residential users. Its operational reliability directly affects the order of social production and life, as well as the security of energy supply. Medium and low voltage distribution networks generally exhibit complex topologies, numerous branch nodes, and frequent load fluctuations. Furthermore, with the large-scale grid connection of distributed power sources and the widespread integration of nonlinear loads such as electric vehicle charging stations, the causes of faults, the amplitude and direction of fault currents, and the manifestation of fault characteristics have all changed significantly, posing a severe challenge to the rapid and accurate identification of faults.

[0003] However, traditional fault identification methods often rely on simple algorithms, failing to fully characterize the coupling relationship between temporal abrupt changes in faults and spatial topology propagation. They cannot fully utilize the time-frequency information in power grid operation data, nor can they effectively mine the spatial correlation characteristics of the power grid topology. Furthermore, current methods struggle to comprehensively and meticulously model the influencing factors of distribution network equipment faults, making it difficult to meet the high-precision fault identification requirements under tree-like topologies and complex operating conditions in distribution networks. Summary of the Invention

[0004] This application provides a fault identification method and related apparatus for a power distribution network, which improves the accuracy of fault identification while reducing the demand for computing power.

[0005] In a first aspect, embodiments of this application provide a method for fault identification in a power distribution network, including: Obtain the first output result of the reference spatiotemporal convolution model, which includes M spatiotemporal convolutional layers in series, each spatiotemporal convolutional layer being used to output spatiotemporal features of different feature dimensions; Determine the target feature dimension of the first output result; Based on the target feature dimension, the feature output dimension of the reference spatiotemporal convolution model is adjusted to obtain the target spatiotemporal convolution model; The dynamic data is input into the target spatiotemporal convolutional model, and the target spatiotemporal features are output. The dynamic data includes the electrical operating status of the power distribution network and the external environmental status. The target spatiotemporal features and static data are fused to obtain initial fault features. The static data includes the topology, electrical parameters and operation and maintenance records of the distribution network. Determine the first contribution of each feature in the initial fault features; The initial fault features are filtered based on the first contribution to obtain the target fault features; Fault identification is performed based on the target fault characteristics to obtain fault results.

[0006] Wherein, determining the target feature dimension of the first output result includes: Determine the second contribution of each feature in the first output result; Determine the first time and the first recognition accuracy for performing the fault identification task based on the first output result; The target feature dimension of the first output result is determined based on the first recognition accuracy, the first time, and the second contribution.

[0007] The step of determining the target feature dimension of the first output result based on the first recognition accuracy, the first time, and the second contribution includes: Obtain the second output result of the (M-1)th spatiotemporal convolutional layer in the reference spatiotemporal convolutional model; Determine the second time and the second identification accuracy for performing the fault identification task based on the second output result; The decrease in the first recognition accuracy is determined based on the second recognition accuracy. The reduction rate of the first time is determined based on the second time. Determine the stability of the second contribution across multiple decision trees; The target feature dimension is determined based on the decrease magnitude, the reduction magnitude, and the stability level.

[0008] The step of determining the target feature dimension based on the decrease magnitude, the reduction magnitude, and the stability degree includes: If the detected decrease is greater than the first threshold, the feature dimension of the first output result is expanded to obtain the target feature dimension; If the stability level is detected to be greater than the second threshold, the first output result is filtered according to the second contribution to obtain the target feature dimension; If the reduction is detected to be less than the third threshold, the feature dimension of the first output result is compressed to obtain the target feature dimension.

[0009] The step of determining the target feature dimension based on the decrease magnitude, the reduction magnitude, and the stability degree includes: Determine the impact of the decrease, reduction, and stability on the fault identification task, respectively. Determine the difference between the decrease magnitude, the reduction magnitude, and the stability level, respectively, and the corresponding threshold values; Based on the difference and the degree of impact, the priorities of the decrease magnitude, the reduction magnitude, and the stability level are determined; The target feature dimension is determined based on the priority.

[0010] The step of inputting dynamic data into the target spatiotemporal convolutional model and outputting target spatiotemporal features includes: The dynamic data is input into the first spatiotemporal convolutional layer, and the first spatiotemporal feature is output. The first spatiotemporal feature is input into the second spatiotemporal convolutional layer, and the second spatiotemporal feature is output. The dimension of the second spatiotemporal feature is smaller than the dimension of the first spatiotemporal feature. The second spatiotemporal feature is input into the third spatiotemporal convolutional layer, and the third spatiotemporal feature is output. The dimension of the third spatiotemporal feature is smaller than that of the second spatiotemporal feature. The third spatiotemporal feature is input into the fourth spatiotemporal convolutional layer, and the target spatiotemporal feature is output. The dimension of the target spatiotemporal feature is smaller than the dimension of the third spatiotemporal feature.

[0011] Each spatiotemporal convolutional layer includes a temporal convolutional layer and a spatial convolutional layer. The step of inputting the dynamic data into the first spatiotemporal convolutional layer and outputting the first spatiotemporal feature includes: The state change characteristics of each node in the power distribution network within different time windows are extracted through the temporal convolutional layer in the first spatiotemporal convolutional layer. Determine the spatial association strength between nodes to obtain the weighted adjacency matrix; The weighted adjacency matrix and the state change features are fused through the spatial convolution layer in the first spatiotemporal convolution layer to obtain the first spatiotemporal feature.

[0012] Secondly, embodiments of this application provide a fault identification device for a power distribution network, comprising: The acquisition unit is used to acquire the first output result of the reference spatiotemporal convolution model, which includes M spatiotemporal convolutional layers in series, each spatiotemporal convolutional layer being used to output spatiotemporal features of different feature dimensions; The first determining unit is used to determine the target feature dimension of the first output result; The adjustment unit is used to adjust the feature output dimension of the reference spatiotemporal convolution model according to the target feature dimension to obtain the target spatiotemporal convolution model. The feature extraction unit is used to input dynamic data into the target spatiotemporal convolution model and output the target spatiotemporal features. The dynamic data includes the electrical operating status of the power distribution network and the external environment status. The fusion unit is used to fuse the target spatiotemporal features and static data to obtain initial fault features. The static data includes the topology, electrical parameters and operation and maintenance records of the distribution network. The second determining unit is used to determine the first contribution of each feature in the initial fault features; A filtering unit is used to filter the initial fault features based on the first contribution to obtain target fault features; The identification unit is used to identify faults based on the target fault characteristics and obtain fault results.

[0013] Thirdly, embodiments of this application provide an electronic device, including a memory, a processor, and executable program code stored in the memory and executable on the processor, wherein the processor executes the executable program code and performs the steps of the method described in the first aspect.

[0014] Fourthly, embodiments of this application provide a computer-readable storage medium storing executable program code, the executable program code including execution instructions for performing the steps of the method as described in the first aspect.

[0015] Fifthly, embodiments of this application provide a computer program product, wherein the computer program product includes a non-transitory computer-readable storage medium storing a computer program operable to cause a computer to perform some or all of the steps described in the first aspect of embodiments of this application. The computer program product may be a software installation package.

[0016] As can be seen, in this embodiment, the first output result of a reference spatiotemporal convolutional model is first obtained. The reference spatiotemporal convolutional model includes M cascaded spatiotemporal convolutional layers, each used to output spatiotemporal features of different feature dimensions. Then, the target feature dimension of the first output result is determined. Next, the feature output dimension of the reference spatiotemporal convolutional model is adjusted according to the target feature dimension to obtain a target spatiotemporal convolutional model. Then, dynamic data is input into the target spatiotemporal convolutional model to output target spatiotemporal features. The dynamic data includes the electrical operation status of the distribution network and the external environment status. Then, the target spatiotemporal features and static data are fused to obtain initial fault features. The static data includes the topology, electrical parameters, and operation and maintenance records of the distribution network. Then, the first contribution of each feature in the initial fault features is determined. Then, the initial fault features are filtered according to the first contribution to obtain target fault features. Finally, fault identification is performed based on the target fault features to obtain fault results.

[0017] This application enhances the ability to deeply represent the temporal abrupt changes and spatial topology propagation of faults through multi-layer spatiotemporal convolution, and adaptively optimizes the feature output dimension of the model to achieve accurate extraction of spatiotemporal coupling features of distribution network faults. It integrates the spatiotemporal features output by the model with static data in multiple dimensions, and eliminates redundant features through feature contribution filtering to reduce computational consumption. Finally, it achieves accurate identification of distribution network faults through the filtered fault features, significantly improving the robustness and generalization ability of fault identification under complex operating conditions. While effectively improving fault identification accuracy, it reduces the demand for computing power and lowers the hardware deployment cost and maintenance difficulty of the terminal. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This application provides a system architecture diagram of a fault identification system. Figure 2 This is a flowchart illustrating a fault identification method for a power distribution network provided in an embodiment of this application; Figure 3 This is a flowchart illustrating a method for determining the target feature dimension of a first output result according to an embodiment of this application; Figure 4 This is a flowchart illustrating another fault identification method for a power distribution network provided in an embodiment of this application; Figure 5 This is a flowchart illustrating another fault identification method for a power distribution network provided in an embodiment of this application; Figure 6 This is a functional unit block diagram of a fault identification device for a power distribution network provided in an embodiment of this application; Figure 7 This is a functional unit block diagram of another fault identification device for a power distribution network provided in this application embodiment; Figure 8 This is a schematic diagram of the structure of an electronic device proposed in an embodiment of this application. Detailed Implementation

[0020] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0021] The terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.

[0022] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0023] As the core terminal link of the power system, the distribution network plays a crucial role in distributing electrical energy from the transmission network to industrial, commercial, and residential users. Its operational reliability directly affects the order of social production and life, as well as the security of energy supply. Medium and low voltage distribution networks generally exhibit complex topologies, numerous branch nodes, and frequent load fluctuations. Furthermore, with the large-scale grid connection of distributed power sources and the widespread integration of nonlinear loads such as electric vehicle charging stations, the causes of faults, the amplitude and direction of fault currents, and the manifestation of fault characteristics have all changed significantly, posing a severe challenge to the rapid and accurate identification of faults.

[0024] However, traditional fault identification methods often rely on simple algorithms, failing to fully characterize the coupling relationship between temporal abrupt changes in faults and spatial topology propagation. They cannot fully utilize the time-frequency information in power grid operation data, nor can they effectively mine the spatial correlation characteristics of the power grid topology. Furthermore, existing data-driven models only consider the relationship between various influencing factors and distribution network faults at each time segment, neglecting the cumulative effect of some factors on distribution network faults. This makes it difficult to comprehensively and meticulously model the influencing factors of distribution network equipment faults, resulting in an inability to meet the high-precision fault identification requirements under tree-like topologies and complex operating conditions in distribution networks.

[0025] In addition, traditional methods require a large number of samples for calculation, which places high demands on computing resources and significantly increases the system operating cost, further limiting their application in actual power distribution networks.

[0026] To address the aforementioned problems, this application provides a fault identification method and related apparatus for power distribution networks. The embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0027] Please see Figure 1 , Figure 1 This is a system architecture diagram of a fault identification system provided in an embodiment of this application. Figure 1 As shown, the fault identification system 100 includes an edge cluster 101, a distribution network master station 102, and a mobile terminal 103, which are interconnected.

[0028] The edge cluster 101 includes a local server cluster and an edge cloud component, which are used to collect data from various systems or terminals, complete preliminary integration and temporary storage, and then transmit the data to the distribution network master station 102.

[0029] Specifically, transparent terminals are deployed at each monitoring node of the distribution network to ensure that at least one monitoring point is set up in each branch of the distribution network. The transparent terminal is the on-site operation monitoring terminal of the distribution network, which is used to collect operating status data such as equipment voltage and current.

[0030] For example, the transparent terminal can use a sampling frequency of 4KHz to collect electrical parameter data such as voltage amplitude, phase angle, active power, and reactive power of each node in the distribution network. The continuous sampling data is encapsulated into sample data every 100ms. These real-time collected electrical parameter data will be directly uploaded to the edge cluster 101, which will complete the preliminary data verification and temporary storage. At the same time, historical data of the terminal for more than one year and which has been preliminarily cleaned by the edge computing nodes will be selected to form the original training dataset for model training.

[0031] It should be noted that the sampling frequency, data recording interval, and other parameters mentioned above are for illustrative purposes only. In practical applications, they can be adjusted according to the topological complexity of the distribution network, the equipment monitoring requirements, and the computing power capacity of the edge computing nodes.

[0032] Specifically, environmental monitoring terminals are deployed on high towers or tall buildings near key nodes of the distribution network to collect lightning strike data in real time, add timestamps, and upload it to the edge cluster 101. Temperature, humidity, and rainfall information can be collected every 30 minutes, simultaneously with timestamps, and transmitted to the edge cluster 101. The edge cluster 101 then performs preliminary correlation and integration with the corresponding electrical parameter data before uniformly transmitting it to the distribution network master station 102. The environmental data collection cycle described here is also an example and can be flexibly adjusted based on regional climate characteristics and the monitoring focus on fault causes.

[0033] Among them, the edge cluster 101 is also connected to the power distribution network geographic information system, which is used to export the power distribution network line topology map and line attribute table, including information such as node location, line type, length, and impedance; at the same time, it is connected to the production management system, which is used to export equipment commissioning records and historical fault records, such as the number of faults and fault types of a certain line in the past 5 years.

[0034] In this process, the edge cluster 101 preprocesses the collected data and then transmits it to the distribution network master station 102. The distribution network master station 102 uses the data to identify faults, such as perform fault prediction and fault diagnosis, and generates corresponding fault prediction reports and fault diagnosis reports.

[0035] Among them, the mobile terminal 103 is a handheld device for operation and maintenance personnel. The distribution network master station 102 pushes information such as fault notification, fault prediction report, and fault diagnosis report to the mobile terminal 103 in real time. The operation and maintenance personnel receive the above information through the mobile terminal 103 and determine the final handling plan based on the actual working conditions on site.

[0036] Based on this, this application provides a fault identification method for power distribution networks, which will be described in detail below with reference to the accompanying drawings.

[0037] Please see Figure 2 , Figure 2 This is a flowchart illustrating a fault identification method for a power distribution network provided in an embodiment of this application, as shown below. Figure 2 As shown, the method includes the following steps: S210, obtain the first output result of the reference spatiotemporal convolution model.

[0038] The first output result is the output result of the Mth spatiotemporal convolutional layer in the reference spatiotemporal convolutional model.

[0039] Among them, the reference spatiotemporal convolution model is a pre-trained deep learning model used to fuse spatial topological features and time series features to output spatiotemporal features.

[0040] The reference spatiotemporal convolution model includes M cascaded spatiotemporal convolutional layers. Each spatiotemporal convolutional layer includes a temporal convolutional layer and a spatial convolutional layer, and each spatiotemporal convolutional layer outputs spatiotemporal features with different feature dimensions.

[0041] The temporal convolutional layer can be a temporal feature extraction module such as a one-dimensional convolutional neural network (1D-CNN), a temporal convolutional network (TCN), or a gated recurrent unit (GRU) to extract features from the temporal data of each node. 1D-CNN and TCN rely on convolutional operations to mine local temporal correlation features of the data, while GRU uses a gating mechanism to capture long-term temporal dependencies, thereby extracting the trend of data changes over time, such as the sudden current changes during a fault.

[0042] Among them, the spatial convolutional layer can be a graph convolutional neural network (GCN), which abstracts the nodes and lines of the distribution network as nodes and edges of a graph, and constructs the relationship between nodes based on the topological adjacency matrix of the distribution network. Then, the fault features of adjacent nodes are aggregated through graph convolution operation, the spatial relationship features between nodes are mined, the topological propagation relationship of electrical parameters of different monitoring nodes is captured, and finally the spatial diffusion law of faults in the power grid is reflected.

[0043] Among them, single-layer TCN and GCN are difficult to capture complex spatiotemporal coupling patterns, such as the propagation of multi-node cascading failures. Therefore, dimensionality compression can be achieved by stacking multiple layers of spatiotemporal convolution and adding pooling operations. For example, if the input is 16-dimensional, the first layer output is 8-dimensional, the second layer output is 6-dimensional, and the third layer output is 4-dimensional.

[0044] In one possible implementation, an initial spatiotemporal convolutional model is first constructed, building the STGCN model framework, which includes M cascaded spatiotemporal convolutional layers. Simultaneously, multi-dimensional data under normal operation and various fault conditions of the distribution network are collected to form a sample set. For example, this includes eight features: three-phase voltage, current, active and reactive power, temperature and humidity, rainfall, and whether lightning strikes occurred. The sample set is then preprocessed, dividing it into a training set, a validation set, and a test set. The initial spatiotemporal convolutional model is iteratively trained based on the training set. The model's hyperparameters, such as learning rate, kernel size, and feature dimension configuration of the M spatiotemporal convolutional layers, are adjusted in real-time using the validation set. The model's recognition accuracy and generalization ability are verified using the test set until the model converges and meets preset performance indicators, resulting in a trained reference spatiotemporal convolutional model, which is then deployed in actual distribution network fault monitoring scenarios.

[0045] Specifically, the trained model is deployed to the power distribution network master station or edge computing node, connected to the terminal data collected in real time, and the output results of the model are obtained, namely the multi-dimensional feature set output by the last spatiotemporal convolutional layer.

[0046] S220, Determine the target feature dimension of the first output result.

[0047] Different scenarios or nodes have different requirements for the dimension of feature data when performing fault identification. The feature output dimension can be adaptively optimized according to the specific power distribution network scenario, such as the number of nodes and the distribution of fault types. This avoids information loss caused by excessive compression and prevents dimensional redundancy from affecting computational efficiency, ensuring the accuracy of feature extraction and engineering practicality.

[0048] In one possible embodiment, please refer to Figure 3 , Figure 3 This is a flowchart illustrating a method for determining the target feature dimension of a first output result according to an embodiment of this application. Figure 3 As shown, it includes the following steps: S310, determine the second contribution of each feature in the first output result.

[0049] The first output is the multi-dimensional spatiotemporal features output from the Mth spatiotemporal convolutional layer of the reference spatiotemporal convolutional model. The importance score, or second contribution, of each feature can be calculated using a decision tree algorithm. The decision tree measures the discriminative power of a feature by calculating the decrease in its information gain or Gini coefficient at the task decision node; the greater the decrease in information gain or Gini coefficient, the greater the contribution of that feature to the determination of the fault state.

[0050] S320, determine the first time and the first recognition accuracy for performing the fault identification task based on the first output result.

[0051] The fault identification task includes fault diagnosis and fault prediction tasks. The performance and efficiency metrics of the reference spatiotemporal convolutional model in these tasks are evaluated, namely, identification accuracy and model inference time. Specifically, the matching rate between the model output and the true labels is calculated for both tasks to obtain the first identification accuracy; the model inference time for fault prediction and fault diagnosis tasks is recorded. The first time refers to the model inference time, which is the time it takes for the model to generate an identification result from the input sample data.

[0052] In one possible embodiment, the first identification accuracy can be the identification accuracy of the fault diagnosis task, the identification accuracy of the fault prediction task, or the average of the identification accuracies of the fault diagnosis task and the fault prediction task; the same applies to the first time.

[0053] In one possible embodiment, the first output result within a preset time period can be obtained. Then, the recognition accuracy and model inference time corresponding to each first output result can be determined. The average recognition accuracy and average model inference time of all output results are calculated to obtain the final first time and first recognition accuracy. Specifically, for each feature set corresponding to each first output result within the preset time period, the importance score of each feature in the feature set is calculated based on multiple decision trees. Subsequently, the average importance score of the same feature in all first output results is calculated, and this average is used as the final second contribution of that feature.

[0054] S330, determine the target feature dimension of the first output result based on the first recognition accuracy, the first time, and the second contribution.

[0055] Different adjustment schemes are set for different evaluation indicators. Specifically, if the accuracy of the first identification is low, the feature dimensions do not meet the needs of the scenario. This may be due to excessive compression of the dimensions, resulting in the loss of key information. In this case, it is necessary to expand the feature output dimensions, retain more temporal details, and ensure that spatially related features are not lost.

[0056] Specifically, if the second contribution is low, the current output features are not discriminative enough for the fault identification task. This may be due to feature omission or an excessive proportion of redundant features. It is necessary to recalculate the importance score of all features based on the decision tree algorithm to obtain spatiotemporal correlation features with high contribution, while removing redundant features with low contribution.

[0057] Specifically, if the initial time is long, the model inference efficiency will not meet the real-time requirements of the distribution network, and the feature dimensions need to be further compressed. At the same time, the spatiotemporal convolution model and / or fault identification model can be optimized to reduce the model inference time while ensuring that the identification accuracy is basically stable.

[0058] In one possible embodiment, determining the target feature dimension of the first output result based on the first recognition accuracy, the first time, and the second contribution includes: obtaining the second output result of the (M-1)th spatiotemporal convolutional layer in the reference spatiotemporal convolutional model; determining the second time and the second recognition accuracy for performing the fault recognition task based on the second output result; determining the decrease magnitude of the first recognition accuracy based on the second recognition accuracy; determining the decrease magnitude of the first time based on the second time; determining the stability of the second contribution among multiple decision trees; and determining the target feature dimension based on the decrease magnitude, the decrease magnitude, and the stability.

[0059] Specifically, the performance loss of the fault identification task is evaluated when the feature dimension is compressed from the (M-1)th layer to the Mth layer, to determine whether key features are lost during dimensionality reduction. This involves comparing the second output of the (M-1)th spatiotemporal convolutional layer with the first output of the Mth spatiotemporal convolutional layer to determine whether the output dimension needs to be optimized.

[0060] Specifically, the second output result is input into the fault prediction model or fault diagnosis model, and the second time and second identification accuracy of the second output result in performing the fault prediction task or fault diagnosis task are calculated.

[0061] The decrease in recognition accuracy is calculated as (first recognition accuracy - second recognition accuracy) / first recognition accuracy × 100%. If the decrease exceeds the preset threshold, it indicates that the output of the Mth layer has lost key spatiotemporal features, and the dimensionality reduction has too great an impact on performance.

[0062] In one possible embodiment, the fault diagnosis task is more sensitive to accuracy and needs to accurately identify the fault type, while the fault prediction task has a relatively higher tolerance for small performance losses, and the degradation threshold of the fault diagnosis task can be set relatively small.

[0063] The reduction in inference time is calculated as (first time - second time) / first time × 100%. If the reduction is less than the preset threshold, it indicates that the feature compression is incomplete, resulting in both performance loss and no efficiency improvement.

[0064] In one possible embodiment, the fault prediction task has a larger input sample size and typically takes longer to infer, so the required reduction in time for this task can be relatively higher.

[0065] In this process, for each decision tree, the second contribution score of each feature is calculated, resulting in a set of scores. All decision trees in the ensemble model are traversed, collecting the score data for each feature across different decision trees. For example, for feature A, N data samples can be collected, including [tree 1 score, tree 2 score, ... tree N score]. The variance of each feature's N contribution scores is calculated; this variance is the quantitative indicator of the stability of the feature's contribution. If the variance exceeds a preset threshold, it indicates that the feature's discriminative value for fault identification is highly random and unreliable, likely belonging to a redundant feature or a feature affected by noise.

[0066] In one possible embodiment, the variance threshold for feature contribution can also be configured differently according to the needs of the two types of tasks. The variance threshold for the fault diagnosis task can be appropriately reduced compared to the fault prediction task, so as to balance the diversity of features and the generalization ability of the model.

[0067] In one possible embodiment, determining the target feature dimension based on the decrease magnitude, the reduction magnitude, and the stability level includes: if the decrease magnitude is detected to be greater than a first threshold, then expanding the feature dimension of the first output result to obtain the target feature dimension; if the stability level is detected to be greater than a second threshold, then filtering the first output result according to the second contribution to obtain the target feature dimension; if the reduction magnitude is detected to be less than a third threshold, then compressing the feature dimension of the first output result to obtain the target feature dimension.

[0068] If the decrease is greater than the first threshold, the feature dimension of the output of the Mth layer is directly expanded, and the currently compressed feature dimension is extended to a higher-dimensional baseline dimension, retaining more deep spatiotemporal features.

[0069] If the stability level is greater than the second threshold, the feature dimension of the first output result needs to be compressed. Based on the second contribution and the corresponding preset threshold, a preset proportion or a preset number of features are retained. Alternatively, the number of spatiotemporal convolutional layers can be increased to achieve dimensionality compression.

[0070] If the reduction is less than the third threshold, the feature dimension of the first output result needs to be compressed. This can be done by performing model lightweighting or directly reducing the feature dimension of the model output.

[0071] For example, the first threshold can be 5%. If M is 4, the feature dimension of the 4th layer output is 4, and the feature dimension of the 3rd layer output is 8. The proportion of correctly predicted or diagnosed fault samples to the total number of fault samples is calculated. The diagnostic accuracy corresponding to the 4-dimensional feature should decrease by less than or equal to 5% compared to the 8-dimensional feature; otherwise, it indicates excessive dimensional compression and loss of key information. If the accuracy decreases by more than 5%, the output dimension of the 4th layer TCN can be increased to 6, i.e., the output dimension is expanded from 4 to 6, while the convolution kernel size remains at 3 to retain more temporal details. The core adjacency matrix of the GCN module is expanded to include more far-end node paths to ensure that spatial correlation features are not lost. The metrics are recalculated until the accuracy decreases by less than or equal to 5%.

[0072] For example, the second threshold can be 0.01. If the importance score of a certain feature dimension is less than 0.05 and the variance is greater than 0.01, it indicates that the dimension is redundant and the feature output dimension needs to be adjusted to filter out key features. Specifically, the adjustment method can be to add a feature selection layer after the fourth spatiotemporal convolutional layer, based on the importance score of the random forest, to remove weak features and retain the 3-dimensional core features; if the feature dimension is less than 3 after removal, it is necessary to fall back to the third spatiotemporal convolutional layer, reduce the output dimension from 8 to 5, and then retain the 4-dimensional features through the feature selection layer; recalculate the variance until the variance is less than or equal to 0.01.

[0073] For example, the third threshold can be 30%. The time taken for a single fault identification after inputting 4-dimensional features into the decision tree must be reduced by at least 30% compared to the inference time with 8-dimensional features; otherwise, the efficiency improvement from dimensionality compression is not significant and does not meet the lightweight requirement. A specific adjustment could be to change the global average pooling in the reference spatiotemporal convolutional model to global max pooling, retaining only the maximum value of each feature, i.e., the moment when the fault is most significant. If this still does not meet the requirement, the output dimension of the fourth layer can be compressed to 3, and the decision tree structure can be simplified simultaneously, such as reducing the tree depth to 3. The inference time can then be recalculated until the reduction in inference time is greater than or equal to 30%.

[0074] As can be seen, in this embodiment of the application, by introducing a dimensional optimal solution judgment and dynamic adjustment mechanism, the output dimension can be adaptively optimized according to the specific power distribution network scenario. This avoids information loss caused by excessive compression and prevents dimensional redundancy from affecting computational efficiency, thus ensuring the accuracy of feature extraction and engineering practicality.

[0075] In one possible implementation, the above conditions may be triggered simultaneously, such as the decrease being greater than the first threshold and the decrease being less than the third threshold. In this case, the principle of prioritizing accuracy should be followed. First, the feature dimension should be expanded to ensure that the decrease is less than the first threshold. Then, stable features should be selected. Finally, efficiency should be improved by appropriately compressing the dimension, and the compressed feature set should be used as the final target feature dimension.

[0076] In one possible embodiment, determining the target feature dimension based on the decrease magnitude, the reduction magnitude, and the stability level includes: determining the degree of influence of the decrease magnitude, the reduction magnitude, and the stability level on the fault identification task; determining the difference between the decrease magnitude, the reduction magnitude, and the stability level and their corresponding thresholds; determining the priority of the decrease magnitude, the reduction magnitude, and the stability level based on the difference and the degree of influence; and determining the target feature dimension based on the priority.

[0077] Based on the requirements of the fault identification task, the different degrees of influence of the three indicators are determined. For example, the degree of influence of the decrease is greater than the degree of influence of the stability, and the degree of influence of the stability is greater than the degree of influence of the decrease.

[0078] The process involves determining the degree to which each indicator deviates from a preset threshold, measured by the difference between the value and the corresponding threshold. Since the three indicators have different dimensions, the differences need to be normalized to eliminate these dimensional discrepancies.

[0079] The process involves calculating a priority score for each indicator by combining its impact and deviation. Higher scores indicate higher priority, requiring priority processing. The priority score is calculated as: Indicator Impact Weight × Absolute Value of Normalized Difference. Indicators are sorted from highest to lowest priority, with the highest-scoring indicator being the first priority, followed by the second and third. The feature dimensions corresponding to each indicator are then optimized and adjusted sequentially according to their priority.

[0080] In one possible implementation, the feature output dimension of the reference spatiotemporal convolutional model can be adjusted each time according to the adjustment method corresponding to the first priority. After the dimension adjustment is completed, the values ​​of the three indicators need to be recalculated to ensure that all indicators meet the threshold requirements. The feature dimension obtained at this time is the target feature dimension.

[0081] S230, adjust the feature output dimension of the reference spatiotemporal convolution model according to the target feature dimension to obtain the target spatiotemporal convolution model.

[0082] The determined target feature dimensions are then incorporated into the structural adjustment of the reference spatiotemporal convolutional model to obtain a final usable model that adapts to the actual fault identification task requirements.

[0083] S240, input dynamic data into the target spatiotemporal convolution model and output the target spatiotemporal features.

[0084] The dynamic data includes the electrical operating status of the power distribution network and the status of the external environment.

[0085] All collected dynamic data will be integrated and stored in a three-dimensional vector format. ,in, This represents the total number of monitoring nodes in the distribution network. This represents the total number of time steps, which is related to the data recording frequency. The data is collected dynamically, including eight features: three-phase voltage, current, active and reactive power, temperature and humidity, rainfall, and whether there is a lightning strike.

[0086] Among them, can be used The criteria identify outliers in the data. If the outliers occur continuously for no more than 3 hours, they are corrected by interpolation using the following formula: , in, Represents a node At any moment Missing data, and Indicates at time Known data before and after.

[0087] The repaired data can be normalized according to feature type, as shown in the following formula: , in, Represents the normalized eigenvalues; Represents a node At the moment Class data, These correspond to eight features respectively; and For the first The extreme values ​​of the data are dynamically updated through a sliding window to avoid excessive suppression of the current data by historical extreme values.

[0088] The model's fault prediction and diagnosis both require "input time series segment - output label" sample pairs. A sliding window can be used to divide the time series data into fixed-length sample data, the length of which is determined according to task requirements. The formula for calculating the sliding window is:

[0089] In the formula, Indicates the first Sample data for a sliding window, with dimension [missing information]. ; The length is the sliding length; the colon ":" indicates retrieving all data in the current dimension; Indicates the range of time steps.

[0090] The fault prediction task requires transient characteristics of the sample data before the fault, such as the voltage drop lasting for one hour. It can be set to 3600, which can collect 10 sample data per second. By selecting continuous data from the past hour, 3600 sample points can be obtained. These samples contain the transient characteristics of electrical parameters before the fault occurred, which are used by the model to predict the fault risk level in the next 10 minutes.

[0091] The fault diagnosis task requires capturing the transient features of sample data before and after the fault. Since relevant features typically become apparent within 15 seconds before and after the fault occurs, It can be set to 300. With a sampling frequency of 10Hz, 300 sample points can be collected in 30 seconds of continuous monitoring. These samples cover the characteristics of electrical parameter changes before and after the fault occurs, which can be used by the model to quickly determine the specific type of fault.

[0092] The time window length and sampling frequency configurations mentioned above are for illustrative purposes only and can be adjusted according to actual circumstances.

[0093] In one possible embodiment, the step of inputting dynamic data into the target spatiotemporal convolutional model and outputting target spatiotemporal features includes: inputting the dynamic data into a first spatiotemporal convolutional layer and outputting a first spatiotemporal feature; inputting the first spatiotemporal feature into a second spatiotemporal convolutional layer and outputting a second spatiotemporal feature, wherein the dimension of the second spatiotemporal feature is smaller than the dimension of the first spatiotemporal feature; inputting the second spatiotemporal feature into a third spatiotemporal convolutional layer and outputting a third spatiotemporal feature, wherein the dimension of the third spatiotemporal feature is smaller than the dimension of the second spatiotemporal feature; and inputting the third spatiotemporal feature into a fourth spatiotemporal convolutional layer and outputting the target spatiotemporal feature, wherein the dimension of the target spatiotemporal feature is smaller than the dimension of the third spatiotemporal feature.

[0094] Preferably, the target spatiotemporal convolutional model includes four spatiotemporal convolutional layers and a final pooling layer.

[0095] In one possible embodiment, the step of inputting the dynamic data into the first spatiotemporal convolutional layer and outputting the first spatiotemporal feature includes: extracting the state change features of each node in the power distribution network within different time windows through the temporal convolutional layer in the first spatiotemporal convolutional layer; determining the spatial association strength between nodes to obtain a weighted adjacency matrix; and fusing the weighted adjacency matrix and the state change features through the spatial convolutional layer in the first spatiotemporal convolutional layer to obtain the first spatiotemporal feature.

[0096] Each spatiotemporal convolutional layer includes a temporal convolutional layer and a spatial convolutional layer. The temporal convolutional layer is used to extract local dynamic features from long-term temporal sample data of a single node. (The text then abruptly shifts to a seemingly unrelated topic: sliding window samples.) As input, the model outputs the temporal feature vector of each node. , The number of output features for TCN is set by empirical values, and it is necessary to ensure that the feature dimension is appropriate.

[0097] For each node Local temporal features are extracted by scanning 1D convolutional temporal data, as shown in the following formula:

[0098] in, It is a 1D convolution function; The temporal convolution kernel is set to 3. Since the temporal abrupt changes of distribution network faults usually appear in a very short time, the length of the temporal convolution kernel can be set to 3. This can capture local abrupt changes completely while avoiding redundancy caused by an excessively large window. The step size indicates that the system moves one unit for each time step of the scan, preserving temporal details and ensuring that the GCN can subsequently analyze spatial correlations by combining complete temporal patterns. This is a bias term.

[0099] The temporal characteristics of distribution network faults are not perfectly linear. After convolution, a ReLU activation function is added to obtain... The ReLU activation function simulates neuron activation by suppressing negative features and preserving positive features, introducing nonlinearity so that the model can capture complex temporal patterns.

[0100] Distribution networks typically have a tree-like topology, and the connectivity between adjacent nodes directly affects fault propagation paths. Therefore, spatial correlations can be quantified by constructing an adjacency matrix. The spatial propagation intensity of distribution network faults is inversely proportional to line impedance; therefore, the input adjacency matrix of the GCN (Global Distribution Network) uses a weighted adjacency matrix. As shown in the following formula: , For example, taking a certain power distribution network topology of "node 1 (substation), node 2, node 3 and node 4" as an example, node 1 is connected to node 2. Node 2 connects to nodes 1 and 3. Node 3 connects to nodes 2 and 4. Node 4 is only directly connected to node 3, therefore the binary adjacency matrix is: , To avoid eigenvalue explosion during graph convolution, the weighted adjacency matrix needs to be adjusted. Normalization process is performed to obtain Then, the time step of the failure point. Corresponding time series feature matrix with normalized weighted adjacency matrix Combine and aggregate the features of neighboring nodes. As shown in the following formula: , in, The spatial convolution weight matrix, For spatial convolution feature dimensions, Iterative optimization through model training; This is used as a baseline adjustment term to adjust the basis of the spatial convolution result and avoid feature shift.

[0101] The LeakyReLU activation function is added after spatial convolution to obtain the following result. This refers to the first spatiotemporal feature, which avoids gradient vanishing caused by negative feature values.

[0102] In one possible embodiment, the first spatiotemporal convolutional layer extracts short-term temporal abrupt changes of a single node and the spatial effects of directly adjacent nodes, such as a voltage drop within 0.3 seconds and the immediate impact of a node 1 failure on node 2; TCN convolutional kernel Covering short-term timing of 0.3 seconds; step size Output Dimension The dimension is increased to 16 to expand the feature space and capture more local mutation features; GCN uses a direct neighbor connection matrix to output the dimension. .

[0103] The second spatiotemporal convolutional layer, building upon the first layer, captures mid-term temporal and secondary spatial correlations, such as the continuous voltage drop within 5 minutes and the influence of node 1 on node 3 through node 2, while removing 25% of redundant feature data; the TCN convolutional kernel... Covering 0.5 seconds of mid-term timing; step size Output GCN uses a two-level neighbor adjacency matrix, and the output dimension is... .

[0104] The third spatiotemporal convolutional layer captures long-term temporal patterns and global spatial correlations, such as the fault evolution process within one hour and the impact range of the fault on all network nodes, and then removes 33% of redundant features; TCN convolutional kernels Covering long-term timings of 0.7 seconds; step size Output GCN uses a global neighbor adjacency matrix to output the dimension. .

[0105] The fourth spatiotemporal convolutional layer removes the last 50% of redundant features, retaining only core features strongly correlated with fault type or location, such as the current surge intensity of the fault source node or the voltage fluctuation amplitude of the furthest affected node; the TCN convolutional kernel... Focusing on the characteristics of the core moment of the fault; step size Output GCN uses a core path adjacency matrix, and its output dimension is... .

[0106] Specifically, the output of the last spatiotemporal convolutional layer is subjected to global average pooling to compress the long temporal dimension to 1 dimension, obtaining the spatiotemporal feature vector of each node. The features of each node at different times are then integrated into a temporally averaged feature, as shown in the following formula: , in, The spatiotemporal feature matrix is ​​the final output of the model, with dimensions of . N represents the number of nodes, which contain various key data information; This is the final timing length; This is the output tensor after the second-layer GCN activation.

[0107] In this process, the more than 30,000 time-step features of each node after pooling are condensed into a single 4-dimensional vector, reducing the amount of data by more than 99% and significantly reducing the computational load. Furthermore, the retained 4-dimensional features are more strongly correlated with the faults, greatly reducing the diagnostic speed of the decision tree and laying the foundation for subsequent decision tree computation.

[0108] In one possible embodiment, the target spatiotemporal convolutional model may include other numbers of spatiotemporal convolutional layers, such as 3 layers, 5 layers, etc., which can be adapted and adjusted according to the differentiated needs of the fault identification task, the complexity of the power distribution network topology, the real-time requirements, and the computing resource limitations of the deployment environment.

[0109] As can be seen, in the embodiments of this application, the deep representation capability of fault temporal mutation and spatial topology propagation is enhanced by the multi-level serial progressive model structure. It can extract spatiotemporal features from the original dynamic data layer by layer, and gradually transform from shallow basic temporal and spatial correlation features into deep high-value fault correlation features. This effectively removes noise interference from the original data and improves the relevance and representativeness of the features.

[0110] S250, the target spatiotemporal features and static data are fused to obtain initial fault features.

[0111] The static data includes the topology, electrical parameters, and operation and maintenance records of the power distribution network.

[0112] Specifically, static data includes distribution network line topology diagrams and line attribute tables, such as node locations, line models, lengths, impedances, and other information; static data also includes equipment commissioning records and historical fault records, such as the number of faults and fault types of a certain line in the past 5 years.

[0113] Static data can be integrated using digital encoding. Line type voltage levels are represented using binary arrays [low voltage, medium voltage, high voltage], such as medium voltage which can be represented as... The load level is [Residential, Industrial, Commercial]. For example, residential is represented as... The fault types are [short circuit, grounding, open circuit, intermittent grounding, other]; line length, service life, and historical fault count are directly represented using raw data. The collected static information is stored in a two-dimensional matrix format. ,in, This is a static feature dimension, with each row corresponding to the static information of a monitoring point.

[0114] Specifically, the spatiotemporal feature dataset reflecting dynamic faults and the static dataset reflecting static fault causes are concatenated node-by-node to form a complete fault feature description, as shown in the following formula: , in, This is the final feature fusion matrix, i.e., the initial fault features.

[0115] S260, determine the first contribution of each feature in the initial fault features.

[0116] One approach is to use the random forest algorithm, where the reduction in the Gini coefficient measures the contribution of different features to fault classification. First, the feature importance of a single tree is calculated using the following formula: , in, ,in, The Gini impurity of the parent node; child node The sample accounts for the parent node The proportion; child node Gini impurity; For the first Tree usage features The set from which nodes are split.

[0117] Then, the importance of global features is calculated using the following formula: , in, The number of trees in the random forest. The larger the value, the more characteristic it indicates. The higher the contribution to fault prediction or diagnosis.

[0118] S270, the initial fault features are filtered according to the first contribution to obtain the target fault features.

[0119] Among them, the fusion feature matrix As input, fault labels are output, and 100 decision trees are trained. Then, based on the trained random forest, the global importance of all features is calculated. Finally, an importance threshold is set. ,reserve Features, remove Features; ultimately, the high-dimensional fused features are filtered into a low-dimensional key feature matrix. That is, the target fault characteristics.

[0120] Among them, the spatiotemporal features of the target and the static auxiliary features of the distribution network are deeply integrated, and redundant information is eliminated through feature screening to construct a high-quality decision tree input feature set, thus balancing model accuracy and computational efficiency.

[0121] S280, perform fault identification based on the target fault characteristics to obtain fault results.

[0122] Fault identification can include fault prediction and fault diagnosis. For the two tasks of fault prediction and fault diagnosis, an appropriate decision tree ensemble model is selected to output high-precision and interpretable results based on key features, and decision basis is generated simultaneously to support maintenance personnel in rapid handling.

[0123] In the fault prediction task, the output includes the risk level. The risk level can be identified by 0, 1, and 2. 0 corresponds to no risk, which means that the current operating status of the distribution network is stable and no potential faults have been detected. 1 corresponds to low risk, which means that there are slight abnormal fluctuations in the system, but no fault triggering conditions have been formed yet. 2 corresponds to high risk, which means that significant signs of impending faults have appeared and timely measures need to be taken.

[0124] Among them, the fault prediction task can be carried out using the XGBoost model, namely the gradient boosting tree model. Based on the idea of ​​ensemble learning, the algorithm can effectively improve the accuracy and stability of risk level judgment by iteratively training multiple decision trees and fusing the prediction results of each tree.

[0125] Specifically, the objective function of the XGBoost model is: , in, For the loss function, multi-class cross-entropy loss can be used, and its formula is as follows: , in, Independent hot encoding for real labels, for example, correspond ; For predicting categories The probability of.

[0126] in, For regularization terms, the formula is: , in, The penalty coefficient for leaf nodes; This is the weighted penalty coefficient; This represents the number of leaf nodes; The leaf node weights are used to control the complexity of the tree and avoid overfitting. Finally, the predicted risk level value for each node is output. The results, along with the corresponding fault categories, are explained.

[0127] The output of the fault diagnosis task includes the faulty node and its fault type, which can be diagnosed using a random forest model. Specifically, the classification decision is based on the majority voting principle, and its formula is as follows: , in, For indicator functions; This is a set of fault categories and nodes. Finally, it outputs the faulty nodes and their fault types, along with explanations.

[0128] Furthermore, after analysis using the decision tree model, the output is transformed into a fault prediction report and a fault diagnosis report. The fault prediction report includes a list of high-risk nodes, risk factors, and recommended measures; the fault diagnosis report includes the fault node, fault type, and handling plan. Finally, all reports are uploaded to the distribution network master station and pushed to maintenance personnel in real time. On-site personnel determine the final handling plan, ensuring that the handling process is initiated within 5 minutes of a fault occurring, achieving closed-loop operation of the entire system.

[0129] After obtaining the fault results, dimensional compression feedback and adaptive adjustment operations can continue.

[0130] In this embodiment, the decision tree module outputs quantifiable and verifiable fault criteria. This interpretive design meets the compliance requirements of the power industry for traceable fault analysis and verifiable handling solutions. Furthermore, by combining a spatiotemporal convolutional model with a decision tree, the spatiotemporal convolutional network simultaneously captures the topological correlations and temporal patterns of the power grid. Combined with the deep-level judgment of the decision tree, this effectively improves prediction accuracy while significantly reducing the demand for computing power and lowering the cost of the terminal.

[0131] Furthermore, the embodiments of this application cover all-dimensional protection, including pre-warning and post-disaster diagnosis, which can greatly reduce economic losses caused by power outages and achieve a dual improvement in technical and economic value.

[0132] As can be seen, in this embodiment, multi-layer spatiotemporal convolution enhances the ability to deeply represent the temporal abrupt changes and spatial topology propagation of faults, and adaptively optimizes the feature output dimension of the model, thereby achieving accurate extraction of the spatiotemporal coupling features of distribution network faults. The spatiotemporal features output by the model are fused with static data in multiple dimensions, and redundant features are eliminated through feature contribution screening, reducing computing power consumption. Finally, the selected fault features are used to achieve accurate identification of distribution network faults, significantly improving the robustness and generalization ability of fault identification under complex operating conditions. While effectively improving the accuracy of fault identification, it reduces the demand for computing power and lowers the hardware deployment cost and maintenance difficulty of the terminal.

[0133] In one possible embodiment, please refer to Figure 4 , Figure 4 This is a flowchart illustrating another fault identification method for a power distribution network provided in an embodiment of this application, as shown below. Figure 4As shown, the sampling data preprocessing operation is first performed, then the spatiotemporal features are extracted by STGCN, and then the feature fusion and filtering operation is carried out to fuse the spatiotemporal features and the preprocessed static data. After fusion, the feature filtering operation is performed based on the feature importance score to obtain the target fault features.

[0134] Subsequently, the system integrates decision-making using decision trees, performs fault prediction and diagnosis based on target fault characteristics, and outputs risk warning reports and fault diagnosis reports. Finally, the system dynamically optimizes its own process through dimensional compression feedback and adaptive adjustment mechanisms, and sends relevant feedback information back to the spatiotemporal feature extraction stage to achieve dynamic optimization of feature dimensions, thereby ensuring the accuracy and efficiency of fault prediction and diagnosis.

[0135] Specifically, please refer to Figure 5 , Figure 5 This is a flowchart illustrating another fault identification method for a power distribution network provided in an embodiment of this application, as shown below. Figure 5 As shown, in the data preprocessing step, dynamic data and static data are collected. Dynamic data includes electrical data and environmental data, which are then subjected to anomaly identification and interpolation repair, data slicing, and normalization in sequence. Static data includes distribution network line topology diagrams and line attribute tables, equipment commissioning records, and historical fault records. The static data is first converted into digital codes and then further stored as a two-dimensional matrix to complete the preprocessing operations of dynamic and static data.

[0136] In the spatiotemporal feature extraction step, STGCN outputs 4-dimensional spatiotemporal features. Specifically, temporal features are extracted through 1D convolution of TCN, input into an activation function for computation, and then output. Simultaneously, a weighted adjacency matrix is ​​created through GCN, and combined with the temporal features output by the activation function, spatiotemporal features are extracted, input into an activation function for computation, and then output. Specifically, the feature dimension is compressed from the initial 16 dimensions to 12 dimensions, then to 8 dimensions through multiple spatiotemporal convolutional layers, ultimately obtaining 4-dimensional features, thus completing the spatiotemporal feature extraction operation.

[0137] In the feature fusion and filtering step, the extracted spatiotemporal features and preprocessed static data are concatenated, the importance score of each feature is calculated, key features are filtered based on the importance score, and the final fault features are output.

[0138] In the decision tree ensemble decision-making step, fault prediction and fault diagnosis tasks are performed based on fault characteristics. Specifically, for task 1, fault prediction, an XGBoost model is used for analysis and computation, outputting a risk level prediction result; for task 2, fault diagnosis, a random forest model is used for analysis and computation, outputting a fault diagnosis prediction result. After the results of both tasks are generated, corresponding text reports are generated based on these results, thus completing the entire decision tree ensemble decision-making process. Finally, dimensionality compression feedback and adaptive adjustment can be performed to optimize the spatiotemporal convolutional model.

[0139] As can be seen, in this embodiment, the electrical parameter data acquired by dual-channel ADCs can be fused and corrected using a multimodal redundant data processing algorithm, further improving the acquisition accuracy of the sample data; and the time series features are deeply fused with the spatial topology features of the power grid. Relying on the hierarchical architecture of STGCN, time and spatial features are fused to achieve low-dimensional feature extraction from high-dimensional spatiotemporal data, accurately capturing the temporal trends and spatial patterns of faults in the distribution network; finally, based on the filtered key feature values, fault risk prediction and fault diagnosis results are output through decision tree calculation theory. This greatly simplifies the amount of feature data required for decision tree calculation, effectively improves the operating efficiency of edge computing in the distribution network terminal, reduces the CPU performance requirements of the distribution network terminal, saves costs, and provides more reliable data support for power grid dispatching decisions, risk warning, and operation optimization.

[0140] For examples consistent with the above embodiments, please refer to... Figure 6 , Figure 6 This is a functional unit block diagram of a fault identification device for a power distribution network provided in an embodiment of this application, such as... Figure 6As shown, the fault identification device 60 for a power distribution network includes: an acquisition unit 61, used to acquire a first output result of a reference spatiotemporal convolutional model, the reference spatiotemporal convolutional model including M cascaded spatiotemporal convolutional layers, each spatiotemporal convolutional layer being used to output spatiotemporal features of different feature dimensions; a first determination unit 62, used to determine the target feature dimension of the first output result; an adjustment unit 63, used to adjust the feature output dimension of the reference spatiotemporal convolutional model according to the target feature dimension to obtain a target spatiotemporal convolutional model; and a feature extraction unit 64, used to input dynamic data into the target spatiotemporal convolutional model, and output... The system comprises: a target spatiotemporal feature set, wherein the dynamic data includes the electrical operating status of the distribution network and the external environment status; a fusion unit 65, used to fuse the target spatiotemporal feature set and static data to obtain initial fault features, wherein the static data includes the topology, electrical parameters and operation and maintenance records of the distribution network; a second determination unit 66, used to determine the first contribution degree of each feature in the initial fault features; a filtering unit 67, used to filter the initial fault features according to the first contribution degree to obtain target fault features; and an identification unit 68, used to identify faults according to the target fault features to obtain fault results.

[0141] In one possible embodiment, in determining the target feature dimension of the first output result, the first determining unit 62 is specifically configured to: determine a second contribution of each feature in the first output result; determine a first time and a first recognition accuracy for performing a fault identification task based on the first output result; and determine the target feature dimension of the first output result based on the first recognition accuracy, the first time, and the second contribution.

[0142] In one possible embodiment, in determining the target feature dimension of the first output result based on the first recognition accuracy, the first time, and the second contribution, the first determining unit 62 is further configured to: obtain the second output result of the (M-1)th spatiotemporal convolutional layer in the reference spatiotemporal convolutional model; determine the second time and the second recognition accuracy for performing the fault recognition task based on the second output result; determine the decrease magnitude of the first recognition accuracy based on the second recognition accuracy; determine the decrease magnitude of the first time based on the second time; determine the stability of the second contribution in multiple decision trees; and determine the target feature dimension based on the decrease magnitude, the decrease magnitude, and the stability.

[0143] In one possible embodiment, in determining the target feature dimension based on the decrease magnitude, the reduction magnitude, and the stability level, the first determining unit 62 is further configured to: if the decrease magnitude is detected to be greater than a first threshold, expand the feature dimension of the first output result to obtain the target feature dimension; if the stability level is detected to be greater than a second threshold, filter the first output result according to the second contribution to obtain the target feature dimension; if the reduction magnitude is detected to be less than a third threshold, compress the feature dimension of the first output result to obtain the target feature dimension.

[0144] In one possible embodiment, in determining the target feature dimension based on the decrease magnitude, the reduction magnitude, and the stability level, the first determining unit 62 is further configured to: determine the degree of influence of the decrease magnitude, the reduction magnitude, and the stability level on the fault identification task; determine the difference between the decrease magnitude, the reduction magnitude, and the stability level and their corresponding thresholds; determine the priority of the decrease magnitude, the reduction magnitude, and the stability level based on the difference and the degree of influence; and determine the target feature dimension based on the priority.

[0145] In one possible embodiment, in terms of inputting dynamic data into the target spatiotemporal convolutional model and outputting target spatiotemporal features, the feature extraction unit 64 is specifically configured to: input the dynamic data into a first spatiotemporal convolutional layer and output a first spatiotemporal feature; input the first spatiotemporal feature into a second spatiotemporal convolutional layer and output a second spatiotemporal feature, wherein the dimension of the second spatiotemporal feature is smaller than the dimension of the first spatiotemporal feature; input the second spatiotemporal feature into a third spatiotemporal convolutional layer and output a third spatiotemporal feature, wherein the dimension of the third spatiotemporal feature is smaller than the dimension of the second spatiotemporal feature; and input the third spatiotemporal feature into a fourth spatiotemporal convolutional layer and output the target spatiotemporal feature, wherein the dimension of the target spatiotemporal feature is smaller than the dimension of the third spatiotemporal feature.

[0146] In one possible embodiment, each spatiotemporal convolutional layer includes a temporal convolutional layer and a spatial convolutional layer. Specifically, in inputting the dynamic data into the first spatiotemporal convolutional layer and outputting the first spatiotemporal feature, the feature extraction unit 64 is further configured to: extract the state change features of each node in the power distribution network within different time windows through the temporal convolutional layer in the first spatiotemporal convolutional layer; determine the spatial association strength between nodes to obtain a weighted adjacency matrix; and fuse the weighted adjacency matrix and the state change features through the spatial convolutional layer in the first spatiotemporal convolutional layer to obtain the first spatiotemporal feature.

[0147] It is understood that since the method embodiments and the device embodiments are different presentations of the same technical concept, the content of the method embodiment section in this application should be adapted to the device embodiment section in a synchronous manner, and will not be repeated here.

[0148] In the case of using integrated units, please refer to Figure 7 , Figure 7 This is a functional unit block diagram of another fault identification device for a power distribution network provided in this application embodiment, such as... Figure 7 As shown, the fault identification device 60 for the distribution network includes a processing module 602 and a communication module 601. The processing module 602 controls and manages the operation of the fault identification device 60, for example, executing the steps of the acquisition unit 61, the first determination unit 62, the adjustment unit 63, the feature extraction unit 64, the fusion unit 65, the second determination unit 66, the screening unit 67, and the identification unit 68, and / or performing other processes of the technology described herein. The communication module 601 is used for interaction between the fault identification device 60 and other devices.

[0149] Among them, such as Figure 7 As shown, the fault identification device 60 of the distribution network may also include a storage module 603, which is used to store the program code and data of the fault identification device 60 of the distribution network.

[0150] The processing module 602 can be a processor or controller, such as a central processing unit (CPU), a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. The processor can also be a combination that implements computing functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc.

[0151] The communication module 601 can be a transceiver, RF circuit, or communication interface, etc. The storage module 603 can be a memory.

[0152] All relevant content in each scenario involved in the above method embodiments can be referenced from the functional descriptions of the corresponding functional modules, and will not be repeated here. The above-mentioned fault identification device 60 for the power distribution network can perform the above-mentioned... Figure 2The method for fault identification in the distribution network is shown.

[0153] Please see Figure 8 , Figure 8 This is a schematic diagram of the structure of an electronic device proposed in an embodiment of this application, as shown below. Figure 8 As shown, the electronic device 800 includes a processor 810, a memory 820, a communication interface 830, and one or more programs 821. The one or more programs 821 are stored in the memory and configured to be executed by the processor. When the program is executed, it includes some or all of the steps of any power distribution network fault identification method described in the above method embodiments. The processor, memory, and communication interface are interconnected and complete communication between them.

[0154] The memory can be volatile memory such as Dynamic Random Access Memory (DRAM) or non-volatile memory such as a hard disk drive. The memory stores a set of executable program code, and the processor calls the executable program code stored in the memory to execute some or all of the steps of any fault identification method for a power distribution network as described in the above embodiments of the power distribution network fault identification method.

[0155] As can be seen, the electronic device 800 described in this application embodiment first obtains the first output result of a reference spatiotemporal convolutional model, which includes M cascaded spatiotemporal convolutional layers, each used to output spatiotemporal features of different feature dimensions; then, it determines the target feature dimension of the first output result; next, it adjusts the feature output dimension of the reference spatiotemporal convolutional model according to the target feature dimension to obtain a target spatiotemporal convolutional model; then, it inputs dynamic data into the target spatiotemporal convolutional model to output target spatiotemporal features, the dynamic data including the electrical operation status of the distribution network and the external environment status; then, it fuses the target spatiotemporal features with static data to obtain initial fault features, the static data including the topology, electrical parameters, and operation and maintenance records of the distribution network; then, it determines the first contribution of each feature in the initial fault features; then, it filters the initial fault features according to the first contribution to obtain target fault features; finally, it performs fault identification based on the target fault features to obtain fault results.

[0156] This application enhances the ability to deeply represent the temporal abrupt changes and spatial topology propagation of faults through multi-layer spatiotemporal convolution, and adaptively optimizes the feature output dimension of the model to achieve accurate extraction of spatiotemporal coupling features of distribution network faults. It integrates the spatiotemporal features output by the model with static data in multiple dimensions, and eliminates redundant features through feature contribution filtering to reduce computational consumption. Finally, it achieves accurate identification of distribution network faults through the filtered fault features, significantly improving the robustness and generalization ability of fault identification under complex operating conditions. While effectively improving fault identification accuracy, it reduces the demand for computing power and lowers the hardware deployment cost and maintenance difficulty of the terminal.

[0157] This application also provides a computer storage medium storing a computer program for electronic data interchange, which causes a computer to perform some or all of the steps of any of the methods described in the above method embodiments, wherein the computer includes an electronic device.

[0158] This application also provides a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program operable to cause a computer to perform some or all of the steps of any of the methods described in the above method embodiments. The computer program product may be a software installation package, and the computer may include an electronic device.

[0159] It should be noted that, for the sake of simplicity, the aforementioned methods are described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are optional, and the actions and modules involved are not necessarily essential to this application.

[0160] In the above embodiments, the descriptions of each embodiment have their own emphasis. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0161] In the several embodiments provided in this application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical or other forms.

[0162] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0163] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software program module.

[0164] If the integrated unit is implemented as a software program module and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0165] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage device, which may include: a flash drive, a read-only memory, a random access memory, a magnetic disk, or an optical disk, etc.

[0166] The embodiments of this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The above description of the embodiments is only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A fault identification method for a power distribution network, characterized in that, include: Obtain the first output result of the reference spatiotemporal convolution model, which includes M spatiotemporal convolutional layers in series, each spatiotemporal convolutional layer being used to output spatiotemporal features of different feature dimensions; Determine the target feature dimension of the first output result; Based on the target feature dimension, the feature output dimension of the reference spatiotemporal convolution model is adjusted to obtain the target spatiotemporal convolution model; The dynamic data is input into the target spatiotemporal convolutional model, and the target spatiotemporal features are output. The dynamic data includes the electrical operating status of the power distribution network and the external environmental status. The target spatiotemporal features and static data are fused to obtain initial fault features. The static data includes the topology, electrical parameters and operation and maintenance records of the distribution network. Determine the first contribution of each feature in the initial fault features; The initial fault features are filtered based on the first contribution to obtain the target fault features; Fault identification is performed based on the target fault characteristics to obtain fault results.

2. The method according to claim 1, characterized in that, Determining the target feature dimension of the first output result includes: Determine the second contribution of each feature in the first output result; Determine the first time and the first recognition accuracy for performing the fault identification task based on the first output result; The target feature dimension of the first output result is determined based on the first recognition accuracy, the first time, and the second contribution.

3. The method according to claim 2, characterized in that, The step of determining the target feature dimension of the first output result based on the first recognition accuracy, the first time, and the second contribution includes: Obtain the second output result of the (M-1)th spatiotemporal convolutional layer in the reference spatiotemporal convolutional model; Determine the second time and the second identification accuracy for performing the fault identification task based on the second output result; The decrease in the first recognition accuracy is determined based on the second recognition accuracy. The reduction rate of the first time is determined based on the second time. Determine the stability of the second contribution across multiple decision trees; The target feature dimension is determined based on the decrease magnitude, the reduction magnitude, and the stability level.

4. The method according to claim 3, characterized in that, Determining the target feature dimension based on the decrease magnitude, the reduction magnitude, and the stability degree includes: If the detected decrease is greater than the first threshold, the feature dimension of the first output result is expanded to obtain the target feature dimension; If the stability level is detected to be greater than the second threshold, the first output result is filtered according to the second contribution to obtain the target feature dimension; If the reduction is detected to be less than the third threshold, the feature dimension of the first output result is compressed to obtain the target feature dimension.

5. The method according to claim 3, characterized in that, Determining the target feature dimension based on the decrease magnitude, the reduction magnitude, and the stability degree includes: Determine the impact of the decrease, reduction, and stability on the fault identification task, respectively. Determine the difference between the decrease magnitude, the reduction magnitude, and the stability level, respectively, and the corresponding threshold values; Based on the difference and the degree of impact, the priorities of the decrease magnitude, the reduction magnitude, and the stability level are determined; The target feature dimension is determined based on the priority.

6. The method according to claim 1, characterized in that, The step of inputting dynamic data into the target spatiotemporal convolutional model and outputting target spatiotemporal features includes: The dynamic data is input into the first spatiotemporal convolutional layer, and the first spatiotemporal feature is output. The first spatiotemporal feature is input into the second spatiotemporal convolutional layer, and the second spatiotemporal feature is output. The dimension of the second spatiotemporal feature is smaller than the dimension of the first spatiotemporal feature. The second spatiotemporal feature is input into the third spatiotemporal convolutional layer, and the third spatiotemporal feature is output. The dimension of the third spatiotemporal feature is smaller than that of the second spatiotemporal feature. The third spatiotemporal feature is input into the fourth spatiotemporal convolutional layer, and the target spatiotemporal feature is output. The dimension of the target spatiotemporal feature is smaller than the dimension of the third spatiotemporal feature.

7. The method according to claim 6, characterized in that, Each spatiotemporal convolutional layer includes a temporal convolutional layer and a spatial convolutional layer. The step of inputting the dynamic data into the first spatiotemporal convolutional layer and outputting the first spatiotemporal feature includes: The state change characteristics of each node in the power distribution network within different time windows are extracted through the temporal convolutional layer in the first spatiotemporal convolutional layer. Determine the spatial association strength between nodes to obtain the weighted adjacency matrix; The weighted adjacency matrix and the state change features are fused through the spatial convolution layer in the first spatiotemporal convolution layer to obtain the first spatiotemporal feature.

8. A fault identification device for a power distribution network, characterized in that, include: The acquisition unit is used to acquire the first output result of the reference spatiotemporal convolution model, which includes M spatiotemporal convolutional layers in series, each spatiotemporal convolutional layer being used to output spatiotemporal features of different feature dimensions; The first determining unit is used to determine the target feature dimension of the first output result; The adjustment unit is used to adjust the feature output dimension of the reference spatiotemporal convolution model according to the target feature dimension to obtain the target spatiotemporal convolution model. The feature extraction unit is used to input dynamic data into the target spatiotemporal convolution model and output the target spatiotemporal features. The dynamic data includes the electrical operating status of the power distribution network and the external environment status. The fusion unit is used to fuse the target spatiotemporal features and static data to obtain initial fault features. The static data includes the topology, electrical parameters and operation and maintenance records of the distribution network. The second determining unit is used to determine the first contribution of each feature in the initial fault features; A filtering unit is used to filter the initial fault features based on the first contribution to obtain target fault features; The identification unit is used to identify faults based on the target fault characteristics and obtain fault results.

9. An electronic device, characterized in that, The device includes: The system includes a memory, a processor, and executable program code stored in the memory and executable on the processor, wherein the processor executes the executable program code to perform the steps of the fault identification method for the power distribution network as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores executable program code, which includes execution instructions for performing the steps of the fault identification method for the distribution network as described in any one of claims 1-7.

Images