Diagnosis method and device for fault reason of power distribution network, electronic equipment and storage medium

Abstract

The invention discloses a power distribution network fault cause diagnosis method and device, electronic equipment and a storage medium, and relates to the technical field of power distribution network fault diagnosis or other related fields, and the method comprises the steps: obtaining fault current sequence data of a target power distribution line, the fault current sequence data comprises N pieces of fault current sampling data, n is a positive integer; performing data preprocessing on the fault current sequence data to obtain target time sequence data of a to-be-input model; the target time sequence data is input to a target model, a fault diagnosis result is output, the target model is a pre-trained circulation network model, and the target model is used for analyzing the abnormal discharge process based on the target time sequence data, identifying the waveform form of the abnormal current, determining the fault diagnosis result according to the waveform identification result and outputting the fault diagnosis result. According to the invention, the technical problem of low reliability of the fault diagnosis result due to the fact that the fault reason diagnosis method of the power distribution network often neglects the analysis of the discharge process and the waveform in the prior art is solved.

Description

Technical Field

[0001] This invention relates to the field of power distribution network fault diagnosis technology, and more specifically, to a method and apparatus, electronic equipment, and storage medium for diagnosing the causes of power distribution network faults. Background Technology

[0002] As the terminal network in a power system, the distribution network is primarily responsible for the stable and efficient transmission and distribution of power to end-user equipment, ensuring that end users receive high-quality electricity. Therefore, the stable operation of the distribution network is of great significance to ensuring that end users receive high-quality electricity. However, due to the complex structure and numerous branches of the distribution network, abnormal events occur frequently. Fault diagnosis has become a crucial aspect of distribution network operation and maintenance management, and its accuracy directly affects the safe and stable operation of the distribution network. Traditional distribution network fault diagnosis technology is mainly based on the principle of relay protection, using main protection, backup protection, automatic reclosing, and other devices to achieve fault location and fault type identification. This diagnostic technology has the following drawbacks: 1. Long diagnosis time; 2. Inaccurate fault location; 3. Inability to identify the cause of the fault; 4. Inapplicable to the operation and maintenance management of the distribution network.

[0003] To address the aforementioned shortcomings, with the rapid development of data acquisition and sensing technologies in recent years, sensor technology has been widely applied in power systems, heralding the arrival of an intelligent era for power systems. Sensor technology is the foundation of power system information acquisition. By installing sensors in the power system (e.g., transient waveform fault indicators), real-time information during the power system's operation is collected, providing reliable operational data for the power system. Existing sensor-based fault diagnosis methods mainly rely on various signal analysis methods such as time domain, frequency domain, time-frequency domain, time-series, modal, and time-series-modal analysis to achieve fault diagnosis. However, they often neglect the analysis of abnormal discharge processes and discharge waveforms of fault signals, and do not consider the statistical regularities of these two aspects in single-phase grounding fault scenarios. They ignore the importance of intermediate processes and the statistical regularity of fault cause distribution, resulting in low diagnostic accuracy and large errors compared to actual conditions, making them unsuitable for the operation, maintenance, and management of distribution networks.

[0004] There is currently no effective solution to the above problems. Summary of the Invention

[0005] This invention provides a method, apparatus, electronic device, and storage medium for diagnosing the causes of power distribution network faults, in order to at least solve the technical problem that related methods for diagnosing the causes of power distribution network faults often neglect the analysis of discharge processes and waveforms, resulting in low reliability of fault diagnosis results.

[0006] According to one aspect of the present invention, a method for diagnosing the cause of a power distribution network fault is provided, comprising: acquiring fault current sequence data of a target power distribution line, wherein the fault current sequence data includes N fault current sampling data, where N is a positive integer; performing data preprocessing on the fault current sequence data to obtain target time-series data to be input into a model, wherein the data preprocessing operations include: data alignment, data normalization, and time-series reconstruction; inputting the target time-series data into a target model and outputting a fault diagnosis result, wherein the target model is a pre-trained recurrent network model, the target model being used to analyze the abnormal discharge process and identify the waveform shape of the abnormal current based on the target time-series data, and determining and outputting a fault diagnosis result based on the waveform identification result.

[0007] Further, the step of preprocessing the fault current sequence data includes: obtaining a preset target down-frequency and the sampling frequency of the fault current sequence data, and determining a down-frequency ratio based on the target down-frequency and the sampling frequency; performing data alignment processing on the fault current sequence data based on the down-frequency ratio to obtain a down-frequency sampling sequence, wherein the down-frequency sampling sequence includes: M down-frequency sampling data determined from the N fault current sampling data, where M is a positive integer less than or equal to N.

[0008] Further, the step of preprocessing the fault current sequence data includes: determining the maximum and minimum sampled data in the down-frequency sampling sequence; for each down-frequency sampled data in the down-frequency sampling sequence, calculating a normalized value of the down-frequency sampled data based on the maximum and minimum sampled data to obtain M normalized values; and arranging the M normalized values ​​according to the time sequence in the down-frequency sampling sequence to obtain a normalized sequence.

[0009] Further, the step of preprocessing the fault current sequence data includes: obtaining the total number Q of sampling cycles in the fault current sequence data, where Q is a positive integer; segmenting the normalized sequence into Q segmented subsequences based on the sampling cycles, where each segmented subsequence contains P sequence data, where P is a positive integer; determining the feature data of each sampling cycle based on the segmented subsequences, where each sampling cycle corresponds to a time step of the target model; and arranging the feature data corresponding to the Q sampling cycles according to the time sequence order to obtain the target time sequence data after time sequence reconstruction.

[0010] Further, for each of the sampling cycles, the step of determining the characteristic data of the sampling cycle based on the segmented subsequence includes: for each of the sampling cycles, extracting P sequence data from the segmented subsequence; linearly superimposing the P sequence data to obtain the zero-sequence current time-series data corresponding to the sampling cycle; and determining the zero-sequence current time-series data as the characteristic data of the sampling cycle.

[0011] Furthermore, the target model is composed of the following model layers: a batch normalization layer, used to receive the preprocessed target time series data and perform normalization processing on the target time series data; a recurrent layer, including a first gated recurrent unit, a first random deactivation unit, a second gated recurrent unit, and a second random deactivation unit, used to obtain the normalized target time series data and extract multidimensional time series features from the target time series data; and an output layer, including a batch normalization unit, a fully connected unit, and an activation function unit, used to map the multidimensional time series features to a waveform morphology identification space to obtain the waveform identification result, and determine and output the fault diagnosis result based on the waveform identification result.

[0012] Further, the target model is obtained through the following steps: acquiring a sample dataset and a pre-built initial model, wherein the initial model uses initial model parameters, and the sample dataset records transient current data of S single-phase grounding samples in the power distribution system, where S is a positive integer; dividing the sample dataset according to a preset allocation ratio to obtain a training sample set and a test sample set, and using the training sample set to perform T rounds of iterative updates on the initial model, where T is a first preset value; in each iteration, inputting sample data from the training sample set into the model in R batches, calculating the cross-entropy loss of the current batch and the gradient of the cross-entropy loss of the current batch with respect to each model parameter after each batch input, and determining the model parameter update scheme for the current iteration based on the gradients of all batches, and performing model updates based on the model parameter update scheme, where R is a second preset value; after performing T rounds of iterative updates, using the test sample set to evaluate the accuracy of the latest model, and determining the latest model as the target model if the model accuracy meets the target.

[0013] According to another aspect of the present invention, a diagnostic device for the cause of a power distribution network fault is also provided, comprising: an acquisition unit for acquiring fault current sequence data of a target power distribution line, wherein the fault current sequence data includes N fault current sampling data, where N is a positive integer; a preprocessing unit for performing data preprocessing on the fault current sequence data to obtain target time-series data to be input into a model, wherein the data preprocessing operations include: data alignment, data normalization, and time-series reconstruction; and an input unit for inputting the target time-series data into a target model and outputting a fault diagnosis result, wherein the target model is a pre-trained recurrent network model, the target model being used to analyze the abnormal discharge process and identify the waveform shape of the abnormal current based on the target time-series data, and to determine and output a fault diagnosis result based on the waveform identification result.

[0014] Further, the preprocessing unit includes: a first determining module, used to acquire a preset target down-frequency and the sampling frequency of the fault current sequence data, and determine a down-frequency ratio based on the target down-frequency and the sampling frequency; and a data alignment module, used to perform data alignment processing on the fault current sequence data based on the down-frequency ratio to obtain a down-frequency sampling sequence, wherein the down-frequency sampling sequence includes: M down-frequency sampling data determined from the N fault current sampling data, where M is a positive integer less than or equal to N.

[0015] Furthermore, the preprocessing unit further includes: a second determining module, used to determine the maximum and minimum sampled data in the down-frequency sampling sequence; a first calculation module, used to calculate a normalized value for each down-frequency sampled data in the down-frequency sampling sequence based on the maximum and minimum sampled data, to obtain M normalized values; and a first arranging module, used to arrange the M normalized values ​​according to the temporal order in the down-frequency sampling sequence, to obtain a normalized sequence.

[0016] Furthermore, the preprocessing unit further includes: a first acquisition module, used to acquire the total number Q of sampling cycles in the fault current sequence data, where Q is a positive integer; a data segmentation module, used to segment the normalized sequence into Q segmented subsequences based on the sampling cycles, where each segmented subsequence contains P sequence data, where P is a positive integer; a third determination module, used to determine the feature data of each sampling cycle based on the segmented subsequences, where each sampling cycle corresponds to a time step of the target model; and a second arrangement module, used to arrange the feature data corresponding to the Q sampling cycles according to the time sequence order to obtain the target time sequence data after time sequence reconstruction.

[0017] Further, the third determining module includes: an extraction submodule, used to extract P sequence data from the segmented subsequence for each sampling cycle; a linear superposition submodule, used to linearly superimpose the P sequence data to obtain the zero-sequence current time-series data corresponding to the sampling cycle; and a determining submodule, used to determine the zero-sequence current time-series data as the feature data of the sampling cycle.

[0018] Furthermore, the target model is composed of the following model layers: a batch normalization layer, used to receive the preprocessed target time series data and perform normalization processing on the target time series data; a recurrent layer, including a first gated recurrent unit, a first random deactivation unit, a second gated recurrent unit, and a second random deactivation unit, used to obtain the normalized target time series data and extract multidimensional time series features from the target time series data; and an output layer, including a batch normalization unit, a fully connected unit, and an activation function unit, used to map the multidimensional time series features to a waveform morphology identification space to obtain the waveform identification result, and determine and output the fault diagnosis result based on the waveform identification result.

[0019] Furthermore, the diagnostic device for the cause of power distribution network faults further includes: a second acquisition module, used to acquire a sample dataset and a pre-built initial model, wherein the initial model uses initial model parameters, and the sample dataset records transient waveform current data of S single-phase grounding samples in the power distribution system, where S is a positive integer; an iterative update module, used to divide the sample dataset according to a preset allocation ratio to obtain a training sample set and a test sample set, and use the training sample set to perform T rounds of iterative updates on the initial model, where T is a first preset value; a second calculation module, used to input the sample data from the training sample set into the model in R batches in each iteration, calculate the cross-entropy loss of the current batch and the gradient of the cross-entropy loss of the current batch with respect to each model parameter after each batch is input, and determine the model parameter update scheme for the current iteration based on the gradients of all batches, and perform model updates based on the model parameter update scheme, where R is a second preset value; and a fourth determination module, used to evaluate the accuracy of the latest model using the test sample set after performing T rounds of iterative updates, and determine the latest model as the target model if the model accuracy meets the standard.

[0020] According to another aspect of the present invention, a computer-readable storage medium is also provided, the computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform the method for diagnosing the cause of power distribution network faults as described above.

[0021] According to another aspect of the present invention, an electronic device is also provided, including one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method for diagnosing the causes of power distribution network faults as described in any of the preceding embodiments.

[0022] This invention proposes a method for diagnosing the causes of power distribution network faults. First, fault current sequence data of the target power distribution line is acquired, where each fault current sequence data contains N fault current sampling data points, where N is a positive integer. Then, the fault current sequence data is preprocessed to obtain target time-series data to be input into the model. The data preprocessing operations include data alignment, data normalization, and time-series reconstruction. Finally, the target time-series data is input into the target model, and the fault diagnosis result is output. The target model is a pre-trained recurrent network model used to analyze the abnormal discharge process and identify the waveform of the abnormal current based on the target time-series data. The fault diagnosis result is determined and output based on the waveform identification result.

[0023] In this invention, fault current sequence data of the target distribution line containing rich fault information is obtained as the original information source. Preprocessing operations such as data alignment and normalization reduce the impact of data errors caused by noise, amplitude differences, and sampling frequency differences between different devices in the original data. The time-series data reconstruction step converts long-series data into one-dimensional multi-channel time-series data as model input, alleviating the gradient vanishing problem when recurrent neural networks process long-series data. This allows the model to effectively learn long-term dependencies in the data. Based on this, the model can extract multi-dimensional time-series features from the fault current sequence data, analyze abnormal discharge processes, and identify the waveform morphology of abnormal currents. Finally, the diagnostic results of the fault cause are output based on the waveform morphology. The above process fully utilizes the ability of recurrent neural networks to process sequence data. Establishing a correlation between waveform morphology and fault cause allows for more accurate fault type diagnosis based on the characteristics of the discharge process. This provides a more reliable and intelligent single-phase grounding fault diagnosis scheme for distribution networks, effectively avoiding misdiagnosis caused by relying solely on a single feature or simple threshold mapping. It improves the overall reliability of the fault diagnosis system and solves the technical problem that related technologies often neglect the analysis of discharge processes and waveforms in distribution network fault cause diagnosis methods, resulting in low reliability of fault diagnosis results. Attached Figure Description

[0024] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0025] Figure 1This is a flowchart of an optional method for diagnosing the causes of power distribution network faults according to an embodiment of the present invention;

[0026] Figure 2 This is a flowchart of an optional method for diagnosing the cause of a single-phase grounding fault in a distribution network based on a GRU recurrent neural network, according to an embodiment of the present invention.

[0027] Figure 3 This is a network architecture diagram of an optional GRU recurrent neural network according to an embodiment of the present invention;

[0028] Figure 4 This is a schematic diagram of an optional diagnostic device for the causes of power distribution network faults according to an embodiment of the present invention;

[0029] Figure 5 This is a hardware structure block diagram of an electronic device (or mobile device) for a method of diagnosing the causes of faults in a power distribution network according to an embodiment of the present invention. Detailed Implementation

[0030] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0031] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0032] The following embodiments of the present invention can be applied to various systems / applications / equipment that require power system fault diagnosis and waveform identification of abnormal discharge currents, enabling a fault intelligent analysis mechanism based on a gated recurrent unit (GRU) neural network. This invention utilizes the powerful time-series processing capabilities of the GRU recurrent neural network to perform in-depth feature mining and analysis on the collected single-phase grounding fault current sequences of the distribution network. Through carefully designed preprocessing steps, including data alignment, data normalization, and time-series reconstruction, it ensures that the data input to the model retains its original time-series characteristics while meeting the input requirements of the neural network.

[0033] In practice, after data preprocessing and standardization by the batch normalization layer, the data is fed into the recurrent layer. The core network, consisting of two GRU units and two Dropout layers, plays a crucial role. The GRU units, with their unique update and reset gate mechanisms, can not only capture the long-term dependencies of the current sequence but also effectively alleviate the gradient vanishing problem, ensuring the stability and efficiency of the model when processing long-term time-series data. The introduction of the Dropout layer further enhances the model's generalization ability, enabling it to make accurate diagnoses when faced with new and unseen fault waveforms. The output layer, through a combination of fully connected layers and Softmax layers, transforms the time-series features extracted by the recurrent layer into a class probability distribution, ultimately outputting the classification results of the fault waveform.

[0034] The present invention will now be described in detail with reference to various embodiments.

[0035] Example 1

[0036] According to an embodiment of the present invention, a method for diagnosing the causes of power distribution network faults is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0037] The implementation subject of this invention can be a distribution network intelligent monitoring and diagnostic system, or integrated into various power system monitoring and analysis platforms such as power automation master stations, distribution line monitoring terminals, and intelligent fault detection equipment. By combining GRU recurrent neural network technology, the fault current waveform is morphologically identified and the discharge process is analyzed to accurately infer the specific cause of the fault.

[0038] In the specific implementation process, combining the GRU recurrent neural network technology in deep learning with professional knowledge in the field of power system fault diagnosis, the collected fault current sequence data is first preprocessed, including data alignment, normalization, and time series reconstruction, to ensure that the data meets the input format requirements of the neural network and eliminate the amplitude and time scale differences between different data sources. Then, the preprocessed time series data is input into a specially designed GRU recurrent neural network model. This model efficiently extracts and analyzes the time series features of the waveform through a combination of batch normalization layer, recurrent layer (containing GRU units), and output layer (containing fully connected layer and Softmax layer), thereby realizing intelligent classification of single-phase grounding waveforms.

[0039] By meticulously analyzing the discharge process corresponding to each waveform morphology, the embodiments of the present invention can accurately diagnose the potential causes of faults based on waveform classification results, combined with statistical laws and engineering practices, such as grounding of metal foreign objects, insulator damage, tree-line conflicts, bird nest interference, high-resistance grounding of vegetation, electric shock of small animals, aging and leakage of transformer elbow heads, composite faults or external transient interference, etc.

[0040] The embodiments of the present invention will now be described in detail with reference to the specific implementation steps.

[0041] Figure 1 This is a flowchart of an optional method for diagnosing the causes of power distribution network faults according to an embodiment of the present invention, such as... Figure 1 As shown, the method includes the following steps:

[0042] Step S101: Obtain the fault current sequence data of the target power distribution line, wherein the fault current sequence data contains N fault current sampling data, where N is a positive integer.

[0043] Specifically, the fault current sequence data is the three-phase current data before and after the fault is triggered, recorded by a transient waveform fault indicator installed on the target distribution line. It should be noted that the transient waveform fault indicator is an advanced monitoring device used to capture the transient electrical characteristics of power lines when a fault occurs. In particular, it can record the current changes before and after the fault is triggered with high precision when a single-phase ground fault occurs, providing extremely important data support for subsequent fault analysis and diagnosis.

[0044] When a fault occurs, especially a single-phase ground fault, the current distribution in the distribution network changes significantly, generating a series of complex transient processes, including but not limited to current spikes, oscillations, and attenuation. The fault indicator, through its built-in high-sensitivity current transformer, can capture these subtle current fluctuations and record the transient waveforms of the three-phase current at a high sampling rate, forming fault current sequence data. This data typically includes the normal operating state before the fault trigger point, the sudden current change at the moment of fault triggering, the changes in current characteristics during the fault process, and the current recovery status during the fault recovery phase. The recording of this data is continuous and detailed, covering current changes from several milliseconds before the fault trigger point to a period after fault recovery, providing complete fault process information for subsequent analysis.

[0045] An optional transient waveform fault indicator may also include a timestamp to ensure the timing of fault current sampling data at each sampling point.

[0046] Step S102: Perform data preprocessing on the fault current sequence data to obtain the target time series data to be input into the model. The data preprocessing operations include: data alignment, data normalization, and time series reconstruction.

[0047] It should be noted that data preprocessing is a crucial step in converting the acquired fault current sequence data into a format suitable for model input. This aims to eliminate inconsistencies and noise in the data, enhance its representational capabilities, and thus improve the accuracy of subsequent model training and diagnosis.

[0048] Since there may be various types and sampling frequencies of transient waveform fault indicators installed in the power system, directly merging these data will lead to inaccurate analysis due to inconsistent time baselines. Therefore, all sampled data can be down-aligned according to the same reference frequency. Optionally, step S102 includes: obtaining the preset target down-frequency and the sampling frequency of the fault current sequence data, and determining the down-frequency ratio based on the target down-frequency and the sampling frequency; performing data alignment processing on the fault current sequence data based on the down-frequency ratio to obtain a down-frequency sampling sequence, wherein the down-frequency sampling sequence includes: M down-frequency sampling data determined from N fault current sampling data, where M is a positive integer less than or equal to N.

[0049] Specifically, assuming the target down-frequency is f t Fault current sequence data X s The sampling frequency is f s X s If the sequence length is N, then a down-sampling method can be used to sample X. s The frequency drops to f t Let's assume the down-frequency sampling sequence is X. tThen the m-th data point in the down-frequency sampling sequence can be determined by the following formula: X t (m)=X s (n), where m = 1, 2, ..., M, and M is the down-frequency sampling sequence X. t The sequence length, X s (n) represents the fault current sequence data X s In and down-frequency sampling sequence X t The index of the nearest data point to the m-th data point, where M and n are determined by the following formula: Among them, symbols This indicates the floor function.

[0050] The data alignment process in this embodiment of the invention ensures the comparability of all data samples on the same time scale by unifying the sampling frequency. Specifically, downsampling or interpolation methods can be used to convert high-sampling-rate data to the minimum or standard sampling rate set by the system, thereby aligning the time coordinates of all datasets and providing a foundation for subsequent feature extraction and pattern recognition.

[0051] To further eliminate the impact of inconsistent current data amplitudes on model evaluation weights during model processing, amplitude normalization can be performed on the aligned sampled data. Optionally, step S102 includes: determining the maximum and minimum sampled data in the down-frequency sampling sequence; for each down-frequency sampled data in the down-frequency sampling sequence, calculating the normalized value of the down-frequency sampled data based on the maximum and minimum sampled data to obtain M normalized values; arranging the M normalized values ​​according to the temporal order in the down-frequency sampling sequence to obtain a normalized sequence.

[0052] In the specific implementation process, assume that the down-frequency sampling sequence after data alignment is X t The normalized sequence is represented as but The m-th data point It can be determined by the following formula: Among them, X t (m) represents X t The value of the m-th data point, max(X) t ) and min(X t ) represent sequences X respectively t The maximum and minimum values.

[0053] The data normalization processing in this embodiment of the invention can scale the amplitude range of current data to a standard interval, such as [0, 1] or [-1, 1], thereby eliminating differences in dimensions and amplitude between data. In specific implementation, normalization strategies typically include maximum-minimum normalization, z-score normalization, etc., which ensure that the data, when training the model, is not biased towards the weights of certain features due to the influence of amplitude magnitude, thus improving the model's generalization ability.

[0054] The current data of the power system is time series data, which has obvious periodicity and long-range dependence. In order for the model to capture the characteristics of the time series more effectively, the normalized sampled data can be reconstructed according to time series. Optionally, step S102 includes: obtaining the total number Q of sampling cycles in the fault current sequence data, where Q is a positive integer; dividing the normalized sequence into data segments based on the sampling cycles to obtain Q segmented subsequences, where each segmented subsequence contains P sequence data, where P is a positive integer; for each sampling cycle, determining the feature data of the sampling cycle based on the segmented subsequences, where each sampling cycle corresponds to a time step of the target model; arranging the feature data corresponding to the Q sampling cycles in time series order to obtain the target time series data after time series reconstruction.

[0055] It should be noted that the time series data reconstruction method based on cycle segmentation includes three stages: data segmentation, time step mapping, and feature mapping. Specifically, the transient waveform data after data alignment and normalization is segmented into units of sampling cycles, each cycle corresponds to a time step of the GRU recurrent neural network model, and the zero-sequence current sampling data within each cycle corresponds to the data features of each time step of the model.

[0056] Specifically, let's assume the fault current sequence data after data alignment and normalization is X. raw =[x1,x2,...,x i ,...], where x i Let represent the data at the i-th sampling point of the original data sequence. Then, the reconstructed data sequence can be represented as: in: This represents the reconstructed data sequence at different time steps, where M represents the total number of cycles in the original data sequence. Further, then... With the original data sequence X raw The relationship is as follows:

[0057] Where p represents the number of sampling points per cycle.

[0058] Optionally, for each sampling cycle, the step of determining the characteristic data of the sampling cycle based on the segmented subsequence includes: for each sampling cycle, extracting P sequence data from the segmented subsequence; linearly superimposing the P sequence data to obtain the zero-sequence current time-series data corresponding to the sampling cycle; and determining the zero-sequence current time-series data as the characteristic data of the sampling cycle.

[0059] The timing reconstruction step in this embodiment of the invention converts the original long sequence data into one-dimensional multi-channel timing data suitable for input to recurrent neural networks (such as GRU, LSTM). This typically involves segmenting the data into fixed-length windows (e.g., each window corresponds to one power frequency cycle) and using these windows as the time step sequence input to the model. This ensures that each time step contains current change information over a period of time, which preserves the local features of the time series and facilitates the model's understanding of timing dependencies from a global perspective.

[0060] Step S103: Input the target time series data into the target model and output the fault diagnosis result. The target model is a pre-trained recurrent network model. The target model is used to analyze the abnormal discharge process and identify the waveform shape of the abnormal current based on the target time series data. The fault diagnosis result is determined and output based on the waveform identification result.

[0061] It should be noted that the target time series data (i.e., the fault current sequence data after time series reconstruction) is input into the target model. The pre-trained target model analyzes the waveform morphology and waveform changes during the discharge process based on all zero-sequence current time series data in the target time series data. It obtains the probability distribution of the discharge waveform morphology matching various standard waveform morphologies, and takes the standard waveform morphology with the highest probability as the final waveform identification result. It further determines the fault diagnosis result corresponding to the standard waveform morphology with the highest probability and outputs it in the end.

[0062] Optionally, the target model consists of the following model layers: a batch normalization layer, used to receive preprocessed target time-series data and standardize the target time-series data; a recurrent layer, including a first gated recurrent unit, a first random deactivation unit, a second gated recurrent unit, and a second random deactivation unit, used to acquire standardized target time-series data and extract multidimensional time-series features from the target time-series data; and an output layer, including a batch normalization unit, a fully connected unit, and an activation function unit, used to map multidimensional time-series features to waveform morphology identification space to obtain waveform identification results, and determine and output fault diagnosis results based on the waveform identification results.

[0063] In this embodiment of the invention, a batch normalization layer is introduced before the recurrent layer to receive the preprocessed input data and standardize each feature. This can effectively reduce the internal covariate bias, accelerate the convergence process of the network, and improve the training efficiency and generalization ability of the model.

[0064] Furthermore, the recurrent layer consists of two GRU units (gated recurrent units) and two Dropout layers (random deactivation layers), aiming to effectively extract multi-dimensional temporal features of the recorded waveform data.

[0065] Specifically, the first GRU unit (i.e., the first gated recurrent unit) can be set to receive the output of the batch normalization layer, with 64 hidden state neurons. This GRU unit can effectively capture long-term dependencies through its update gate and reset gate mechanism, while alleviating the gradient vanishing problem in traditional RNN recurrent neural networks. The first Dropout layer (i.e., the first random deactivation unit) is located between the two GRU units, with a dropout rate set to 0.5. This layer can effectively reduce the risk of overfitting and improve the robustness and generalization ability of the model by randomly dropping a certain proportion of neurons during training.

[0066] Similarly, the second GRU unit (i.e., the second gated recurrent unit) has the same structure and function as the first GRU unit, further extracting temporal features and outputting the hidden state of the last time step after processing; the second Dropout layer (i.e., the second random deactivation unit) is set after the second GRU unit, with the same dropout rate as the first Dropout layer, further enhancing the model's generalization ability.

[0067] Furthermore, the output layer maps the temporal features extracted by the recurrent layer to specific waveform morphology categories in the waveform morphology recognition space, thus achieving the final classification decision. The output layer consists of one batch normalization layer, one fully connected layer, and a Softmax layer.

[0068] Specifically, the batch normalization layer is located before the fully connected layer and aims to standardize the output of the recurrent layer, making the feature distribution more stable and helping to stabilize the subsequent classification process. The number of neurons in the fully connected layer is equal to the number of waveform morphology categories to be classified (in this embodiment, a value of 5 is preferred). This fully connected layer receives the 64-dimensional feature vector output by the recurrent layer and maps the feature space to the category recognition space through fully connected transformation. The Softmax layer uses the activation function Softmax to convert the output of the previous layer into a probability distribution. The Softmax function ensures that the output value is in the range [0,1] and the sum of the probabilities of all categories is 1, so that the model output has good statistical interpretation.

[0069] The entire target model selects the waveform category corresponding to the highest probability value based on the probability distribution output by the Softmax layer as the identification result, thereby determining the fault cause corresponding to the input data and obtaining the fault diagnosis result.

[0070] Optionally, the target model is obtained through the following steps: acquiring a sample dataset and a pre-built initial model, wherein the initial model uses initial model parameters, and the sample dataset records transient current data of S single-phase grounding samples in the power distribution system, where S is a positive integer; dividing the sample dataset according to a preset allocation ratio to obtain a training sample set and a test sample set, and using the training sample set to perform T rounds of iterative updates on the initial model, where T is a first preset value; in each round of iteration, inputting sample data from the training sample set into the model in R batches, calculating the cross-entropy loss of the current batch and the gradient of the cross-entropy loss of the current batch with respect to each model parameter after each batch input, and determining the model parameter update scheme for the current iteration based on the gradients of all batches, and performing model updates based on the model parameter update scheme, where R is a second preset value; after performing T rounds of iterative updates, using the test sample set to evaluate the accuracy of the latest model, and if the model accuracy meets the target, determining the latest model as the target model.

[0071] The transient current data includes transient currents and labels derived from the waveform morphology and fault triggering cause analysis of each transient current recorded in the power distribution system fault records. Acquiring the transient current data may also include preprocessing each transient current and its label to obtain target sample data for the input model.

[0072] Specifically, based on practical engineering experience, the classification of zero-sequence current waveforms after a single-phase ground fault in a distribution network in this embodiment of the invention mainly considers the following key characteristics: time domain continuity, including the duration of the waveform, pulse interval, and number of pulses; frequency domain characteristics, considering the harmonic composition and main frequency components of the waveform, reflecting periodicity, spikes, or glitches; and amplitude characteristics, including the peak value and relative amplitude variation of the waveform.

[0073] Based on the above characteristics, the zero-sequence current waveform after a single-phase ground fault in a distribution network can be classified into five typical types.

[0074] 1. Single pulse waveform.

[0075] Specifically, the zero-sequence current waveform of a single pulse exhibits a single, rapid, high-amplitude pulse spike, which then quickly decays to a steady-state level. In terms of waveform shape, it is characterized by only one significant pulse spike occurring within the observation time window, with an extremely short pulse duration, typically on the order of milliseconds. Regarding amplitude characteristics, the spike exhibits a significant amplitude, clearly deviating from the baseline. As for the trend, the current rapidly decreases after the pulse, possibly accompanied by a small amount of oscillation, eventually stabilizing.

[0076] 2. Multi-pulse waveform.

[0077] Specifically, the zero-sequence current waveform of a multi-pulse system exhibits intermittent, multiple pulse-like spikes. In terms of waveform morphology, it consists of multiple pulses, each potentially differing in shape (e.g., rise time, duration, decay rate). Regarding amplitude characteristics, the spikes show significant amplitude, deviating markedly from the baseline. The amplitude of continuous pulses may gradually decay or remain relatively stable. In terms of trend, there are distinct intervals between pulses, each greater than one power frequency cycle (20ms). The multiple pulse spikes show a certain temporal regularity, but are not strictly equal in interval; a brief oscillation or recovery process may occur after each pulse.

[0078] 3. Continuous periodic waveform.

[0079] Specifically, the continuous periodic zero-sequence current waveform exhibits a continuous quasi-sinusoidal periodic change, possibly superimposed with one or two pulses. In terms of waveform morphology, it generally resembles a sine wave, possibly with slight distortion or asymmetry, and the waveform shows obvious periodic changes with a period of 50Hz (power frequency). In terms of amplitude characteristics, the amplitude of the periodic waveform is relatively stable, with little variation in the amplitude of peaks and troughs, possibly superimposed with one or two obvious pulse-like spikes, which are usually large in amplitude. In terms of the trend of change, the duration is greater than two cycles, exhibiting two different characteristics: a) a continuous waveform, maintaining a relatively stable amplitude during the observation period without showing obvious attenuation; b) a rapidly decaying waveform, rapidly decaying to zero after the initial few cycles.

[0080] 4. Spiky waveform.

[0081] Specifically, the spiking zero-sequence current waveform exhibits dense spiking characteristics and lasts for a relatively long time. In terms of waveform morphology, a large number of dense spikes appear in the waveform, showing obvious continuity; in terms of amplitude characteristics, the amplitude of the spikes is relatively small but deviates significantly from the baseline, and the amplitude of each spike may vary, but overall it remains within a certain range; in terms of the trend of change, the waveform lasts for a long time, usually more than one cycle (20ms), without obvious discontinuities or baseline recovery periods, and the spiking characteristics remain relatively stable throughout the observation time, without obvious attenuation trends.

[0082] 5. Other waveforms. All waveforms other than the four types mentioned above are classified into the fifth type of waveform.

[0083] In this embodiment of the invention, a random sampling method can be used to divide the preprocessed sample data into a training set and a test set, with proportions of 80% and 20%, respectively. The training set is used for model parameter learning, while the test set is used to evaluate model performance. Then, an initial model is obtained according to the aforementioned preferred model parameters.

[0084] Alternatively, this embodiment of the invention can also employ batch gradient descent for model training, using 32 samples per batch, with a total of 1000 iterations during the training process, each iteration traversing the entire training set. In each iteration, the cross-entropy loss of the current batch is first calculated, and then the gradient of the loss function with respect to the parameters of each layer is calculated using the backpropagation algorithm.

[0085] Based on this, the momentum method is applied to update the model parameters in this iteration. The calculation formula is as follows: W(t+1) = W(t) - v(t+1), where v(t) and v(t+1) represent the velocity variables before and after the t-th iteration update, respectively, and W(t) and W(t+1) represent the model parameters before and after the t-th iteration update, respectively. The gradient of the loss function with respect to the model parameters is represented by α, which is the momentum coefficient. In this embodiment of the invention, α = 0.9 is preferred, and η is the learning rate. In this embodiment of the invention, η = 0.005 is preferred.

[0086] Furthermore, embodiments of the present invention can use the accuracy on test samples as an evaluation metric for the model, calculated using the following formula: Where, N tes C represents the total number of samples in the test set. tes This indicates the number of samples correctly classified by the model.

[0087] After obtaining the waveform morphology identification result corresponding to the fault current sequence data in step S103, the following mapping rules can be referenced in the process of mapping the fault diagnosis result based on the result.

[0088] For single-pulse waveforms, the discharge process typically manifests as a transient peak discharge, after which the insulation state is restored. Possible causes of the fault include: (1) grounding by a metallic foreign object, such as iron wire or other conductive objects contacting the line; (2) grounding due to insulator damage, such as a porcelain insulator cracking and becoming damp, leading to transient breakdown. In the embodiments of the present invention, the target model preferably outputs "grounding by a metallic foreign object" or "grounding due to insulator damage" as the fault cause diagnosis result.

[0089] For multi-pulse waveforms, the discharge process exhibits intermittent tip discharge, with a brief return to insulation after each discharge. Possible causes of the fault include: (1) tree branches contacting the line, leading to intermittent discharge; (2) bird nest interference, such as intermittent contact between bird nest material and charged parts. In this embodiment of the invention, the target model preferably outputs "tree-line contradiction" or "bird nest interference" as the fault cause diagnosis result.

[0090] For continuous periodic waveforms, the discharge process is characterized by continuous and stable discharge, possibly accompanied by occasional strong discharges. Possible causes of the fault include: (1) dry tree branches forming a stable high-resistance grounding path on the insulated equipment, resulting in a continuous weak current; (2) small animals (such as squirrels, snakes, etc.) being electrocuted, creating a continuous grounding path. In the embodiments of the present invention, the target model preferably outputs "high-resistance grounding of vegetation" or "small animal electrocution" as the fault cause diagnosis result.

[0091] For spiky waveforms, the discharge process is characterized by continuous, weak discharge with a small but frequently fluctuating current amplitude. Possible causes of the fault include: aging of the transformer elbow insulation, resulting in multiple minor leakage points. In this embodiment of the invention, the target model preferably outputs "transformer elbow aging leakage" as the fault cause diagnosis result.

[0092] For other waveforms, the discharge process may exhibit complex or irregular characteristics. Possible causes of failure include: (1) superposition of multiple fault types, such as the simultaneous existence of vegetation disturbance and equipment failure; (2) transient external interference, such as lightning strikes or brief contact with large animals. In the embodiments of the present invention, the target model preferably outputs "composite fault" or "external transient interference" as the fault cause diagnosis result.

[0093] Through the above steps S101 and S103, the fault current sequence data of the target power distribution line can be obtained first. The fault current sequence data contains N fault current sampling data, where N is a positive integer. Then, the fault current sequence data is preprocessed to obtain the target time series data to be input into the model. The data preprocessing operations include data alignment, data normalization, and time series reconstruction. Finally, the target time series data is input into the target model, and the fault diagnosis result is output. The target model is a pre-trained recurrent network model. The target model is used to analyze the abnormal discharge process and identify the waveform shape of the abnormal current based on the target time series data. The fault diagnosis result is determined and output based on the waveform identification result.

[0094] In this embodiment of the invention, fault current sequence data of the target distribution line containing rich fault information is obtained as the original information source. Preprocessing operations such as data alignment and data normalization reduce the impact of data errors caused by noise, amplitude differences, and sampling frequency differences between different devices in the original data. The time-series data reconstruction step converts long-series data into one-dimensional multi-channel time-series data as model input, alleviating the gradient vanishing problem when recurrent neural networks process long-series data. This allows the model to effectively learn long-term dependencies in the data. Based on this, the model can extract multi-dimensional time-series features from the fault current sequence data, analyze the abnormal discharge process, and identify the waveform morphology of the abnormal current. Finally, the diagnostic result of the fault cause is output based on the waveform morphology. The above process fully utilizes the ability of recurrent neural networks to process sequence data. Establishing a correlation between waveform morphology and fault cause allows for more accurate fault type diagnosis based on the characteristics of the discharge process. This provides a more reliable and intelligent single-phase grounding fault diagnosis scheme for distribution networks, effectively avoiding misdiagnosis caused by relying solely on a single feature or simple threshold mapping. It improves the overall reliability of the fault diagnosis system and solves the technical problem that related technologies often neglect the analysis of the discharge process and waveform in distribution network fault cause diagnosis methods, resulting in low reliability of fault diagnosis results.

[0095] The present invention will now be described in conjunction with another specific embodiment.

[0096] This invention proposes a method for diagnosing single-phase grounding faults in distribution networks based on a GRU recurrent neural network. By preprocessing transient waveform data, a GRU recurrent neural network is constructed to identify the waveform morphology of single-phase grounding faults. By analyzing the discharge process corresponding to the morphology, a correlation with the fault cause is established, thereby realizing the identification of the cause of single-phase grounding faults in distribution networks and effectively improving the level of intelligence in fault diagnosis.

[0097] Figure 2 This is a flowchart of an optional method for diagnosing the cause of a single-phase grounding fault in a distribution network based on a GRU recurrent neural network, according to an embodiment of the present invention. Figure 2 As shown, the method includes the following steps:

[0098] S1 classifies the transient waveforms of single-phase grounding in the distribution network and completes the labeling of the samples.

[0099] It should be noted that, compared with three-phase current, zero-sequence current can more directly and sensitively reflect the characteristics of single-phase grounding faults and is not affected by load current. Therefore, the waveform morphology classification method proposed in this invention focuses on the zero-sequence current waveform recorded after a single-phase grounding fault in the distribution network, and can divide the zero-sequence current waveform after a single-phase grounding fault in the distribution network into five typical types.

[0100] Specifically, it includes the following steps:

[0101] S1.1, Determine the basis for waveform morphology classification.

[0102] This invention analyzes a large amount of single-phase grounding transient waveform data and extracts the following key features as the basis for waveform morphology classification: a) time-domain continuity, including waveform duration, pulse interval, and pulse number; b) frequency-domain characteristics, considering the waveform's harmonic composition and main frequency components, reflecting morphological features such as periodicity, spikes, or glitches; c) amplitude characteristics, including the waveform's peak value and relative amplitude variation. Based on these features, a systematic classification standard is established to achieve accurate classification of zero-sequence current waveforms.

[0103] S1.2, Waveform morphology classification.

[0104] According to the waveform morphology classification criteria described in S2.2.1, the implementation of this invention classifies the waveform morphology of a single-phase ground fault in a distribution network into five types.

[0105] S1.2.1, Single-pulse waveform. The zero-sequence current waveform exhibits a single, rapid, high-amplitude pulse spike, which then quickly decays to a steady-state level. Specifically, it is as follows.

[0106] Waveform shape: Only one significant pulse spike appears within the observation time window, and the pulse duration is extremely short, usually on the order of milliseconds;

[0107] Amplitude characteristics: The peaks exhibit significant amplitudes and deviate markedly from the baseline;

[0108] Trend of change: After the pulse, the current drops rapidly, possibly accompanied by a small amount of oscillation, and eventually tends to stabilize.

[0109] S1.2.2, Multi-pulse waveform. The zero-sequence current waveform exhibits intermittent, multiple pulse-like spikes, as shown below.

[0110] Waveform morphology: The waveform contains multiple pulses, and each pulse may differ in morphology (such as rise time, duration, decay rate, etc.).

[0111] Amplitude characteristics: The spikes exhibit significant amplitude, deviating markedly from the baseline. The amplitude of continuous pulses may gradually decay or remain relatively stable;

[0112] Trends: There are distinct intervals between the pulses, each interval being greater than one power frequency cycle (20ms). Multiple pulse spikes exhibit a certain regularity in time, but are not strictly equal in interval. After each pulse, the system may experience a brief oscillation or recovery process.

[0113] S1.2.3, Continuous Periodic Waveform. The zero-sequence current waveform exhibits a continuous quasi-sinusoidal periodic change, possibly superimposed with one or two pulses. Specific manifestations are as follows.

[0114] Waveform shape: The overall shape resembles a sine wave, but slight distortion or asymmetry may exist. The waveform exhibits obvious periodic changes, with a period of 50Hz (power frequency).

[0115] Amplitude characteristics: The amplitude of the periodic waveform is relatively stable, with little variation in the amplitude of the peaks and troughs. It may superimpose one or two distinct pulse-like spikes, which typically have larger amplitudes.

[0116] Trend of change: Duration greater than 2 cycles. It exhibits two different characteristics: (a) a continuous waveform that maintains a relatively stable amplitude during the observation period without showing obvious decay; (b) a rapidly decaying waveform that decays rapidly to zero after the first few cycles.

[0117] S1.2.4, Gibbly Waveform. The zero-sequence current waveform exhibits dense gibbly characteristics and lasts for a relatively long time. Specific manifestations are as follows.

[0118] Waveform morphology: The waveform contains a large number of dense spikes, which have obvious continuity;

[0119] Amplitude characteristics: The amplitude of the burrs is relatively small but deviates significantly from the baseline. The amplitude of each burr may vary, but overall it remains within a certain range.

[0120] Trend of change: The waveform has a long duration, usually more than one cycle (20ms), with no obvious interruption or baseline recovery period. Throughout the observation time, the spiking feature remains relatively stable and there is no obvious attenuation trend.

[0121] S1.2.5, Other waveforms. Waveforms other than the four types mentioned above.

[0122] S1.3, Sample Labeling. Based on the five waveform morphology characteristics defined in S1.2, the collected single-phase grounding transient waveform samples of the distribution network are classified, and each sample is labeled with a corresponding label, thus completing the sample labeling work.

[0123] S2, data preprocessing, transforms the acquired transient waveform data into one-dimensional multi-channel time-series data suitable for recurrent neural networks through data alignment, data normalization, and data reconstruction operations.

[0124] S2.1, Data Alignment: For transient waveform recorders of different models currently installed in the distribution network, their waveform recording frequencies may differ. To enable the waveform recording data from different indicators to be analyzed on the same time scale, this invention proposes a data alignment method based on a downsampling strategy to achieve unified processing of data with different sampling frequencies. The specific implementation scheme is as follows.

[0125] The target frequency f is the lowest sampling frequency. t The sampling data from other devices is downsampled and unified to f. t For a sampling frequency of f s The original data sequence X s Given a sequence of length N, its frequency is reduced to the target frequency f using a downsampling method. t Let the processed data sequence be X. t The m-th data point X t (m) is determined by the following formula:

[0126] X t (m)=X s (n), where M is the downsampled data sequence X. t Length, X s (n) represents the original data sequence X. s In the new sequence X t The index of the nearest data point to the m-th data point, where M and n are determined by the following formula: Among them, symbols This indicates the floor function.

[0127] S2.2, Data Normalization: To eliminate amplitude differences between different data samples and highlight the waveform's morphological characteristics, this invention proposes applying a maximum-minimum value normalization method to process the aligned sample data. The specific implementation scheme is as follows: Let X... t This is the data sequence after data alignment. For the normalized data sequence, then for The value of the m-th data point Determined by the following formula: Among them, X t (m) represents Xt The value of the m-th data point, max(X) t ) and min(X t ) represent sequences X respectively t The maximum and minimum values.

[0128] S2.3, Time Series Data Reconstruction: To adapt to the data format requirements of the GRU recurrent neural network model constructed in this invention and to solve the long-range dependency problem of the GRU network on long-sequence data, this invention proposes a time series data reconstruction method based on cycle segmentation, utilizing the periodic characteristics of transient waveform data. The specific implementation scheme is as follows.

[0129] Data segmentation: The transient waveform data, after data alignment and normalization, is segmented into units of sampling cycles;

[0130] Time step mapping: Each cycle is mapped to a time step in the GRU recurrent neural network model;

[0131] Feature mapping: The zero-sequence current sampling data in each cycle is mapped to the data features of the model at each time step.

[0132] Assume the original transient waveform data sequence after data alignment and normalization is X. raw =[x1,x2,...,x i ,...], where x i Let represent the data at the i-th sampling point of the original data sequence. Then, the reconstructed data sequence can be represented as: in: This represents the reconstructed data sequence at different time steps, where M represents the total number of cycles in the original transient waveform data. With the original data sequence X raw The relationship is as follows: Where p represents the number of sampling points per cycle.

[0133] S3, Design and implement a GRU recurrent neural network for waveform morphology recognition.

[0134] Figure 3 This is a network architecture diagram of an optional GRU recurrent neural network according to an embodiment of the present invention, such as... Figure 3 As shown, the GRU recurrent neural network includes a batch normalization layer, a recurrent layer, and an output layer.

[0135] S3.1, batch normalization layer.

[0136] In this embodiment of the invention, a batch normalization layer is introduced before the recurrent layer. The batch normalization layer receives the preprocessed input data and standardizes each feature. The batch normalization layer effectively reduces internal covariate bias, accelerates the network convergence process, and improves the model's training efficiency and generalization ability.

[0137] S3.2, Loop Layer.

[0138] The cyclic layer constructed in this embodiment of the invention consists of two GRU units and two Dropout layers, aiming to effectively extract multi-dimensional temporal features of the waveform data.

[0139] Specifically, the first GRU unit receives the output of the batch normalization layer, and the number of hidden state neurons is 64. Through its update gate and reset gate mechanism, the GRU unit can effectively capture long-term dependencies and alleviate the gradient vanishing problem in traditional RNN recurrent neural networks.

[0140] A Dropout layer is located between two GRU units, with a dropout rate of 0.5. This layer effectively reduces the risk of overfitting and improves the robustness and generalization ability of the model by randomly dropping a certain proportion of neurons during training.

[0141] The second GRU unit has the same structure as the first GRU unit, and further extracts temporal features. After processing, the hidden state of the last time step is output.

[0142] The second Dropout layer is placed after the second GRU unit, with the same dropout rate as the first Dropout layer, further enhancing the model's generalization ability.

[0143] S3.3, Output Layer.

[0144] The output layer maps the temporal features extracted by the recurrent layer to specific waveform morphology categories, thus achieving the final classification decision. The output layer consists of one batch normalization layer, one fully connected layer, and a softmax layer.

[0145] Specifically, the batch normalization layer, located before the fully connected layer, aims to standardize the output of the recurrent layer, making the feature distribution more stable and contributing to the stability of the subsequent classification process. The number of neurons in the fully connected layer is equal to the number of waveform morphology categories to be classified (preferably 5 in this invention). This layer receives the 64-dimensional feature vector output from the recurrent layer and maps the feature space to the category space through a fully connected transformation. The Softmax layer uses the Softmax activation function to convert the output of the previous layer into a probability distribution. The Softmax function ensures that the output value is within the range [0,1] and that the sum of the probabilities of all categories is 1, giving the model output good statistical interpretation.

[0146] Based on the probability distribution output by the Softmax layer, the model selects the category corresponding to the highest probability value as the prediction result, thereby determining the waveform shape category of the input sample.

[0147] S4, train the GRU recurrent network and evaluate its performance.

[0148] S4.1, Dataset Partitioning: In this embodiment of the invention, a random sampling method is used to divide the preprocessed sample dataset into a training set and a test set, with proportions of 80% and 20%, respectively. The training set is used for model parameter learning, while the test set is used to evaluate the final performance of the model.

[0149] S4.2, Constructing the GRU Recurrent Neural Network: Construct the GRU recurrent neural network according to the model architecture proposed in S3 and initialize the model parameters.

[0150] S4.3, Model Training: This embodiment of the invention employs batch gradient descent for model training, using 32 samples per batch. The training process iterates for 1000 rounds, traversing the entire training set in each round. In each iteration, the cross-entropy loss of the current batch is first calculated, and then the gradient of the loss function with respect to the parameters of each layer is calculated using the backpropagation algorithm. Based on this, the momentum method is applied to update the model parameters for this iteration, and the calculation formula is as follows: W(t+1) = W(t) - v(t+1), where v(t) and v(t+1) represent the velocity variables before and after the t-th iteration update, respectively, and W(t) and W(t+1) represent the parameters of the model before and after the t-th iteration update, respectively; The gradient of the loss function with respect to the model parameters is represented by α, which is the momentum coefficient. In this embodiment of the invention, α = 0.9 is preferred, and η is the learning rate. In this embodiment of the invention, η = 0.005 is preferred.

[0151] S4.4, Performance Evaluation: This invention uses the accuracy on test samples as the evaluation metric for the model, calculated using the following formula: Where, N tes C represents the total number of samples in the test set. tes This indicates the number of samples correctly classified by the model.

[0152] S5 analyzes the discharge process corresponding to the waveform shape and outputs the cause of the single-phase grounding fault.

[0153] This invention analyzes different waveform morphologies and their corresponding discharge processes, and establishes a correlation between waveform characteristics and fault causes based on practical engineering experience, outputting the fault cause with the highest probability. The specific process is as follows.

[0154] S5.1 For single-pulse waveforms, the discharge process is usually characterized by a transient peak discharge, after which the insulation state is restored. Possible causes of the fault include: (1) grounding by a metallic foreign object, such as iron wire or other conductive objects contacting the line; (2) grounding due to insulator damage, such as a porcelain insulator cracking and becoming damp, leading to transient breakdown. In the preferred embodiment of this invention, "grounding by a metallic foreign object" or "grounding due to insulator damage" is output as the fault cause identification result.

[0155] S5.2, for multi-pulse waveforms, the discharge process exhibits intermittent tip discharge, with a brief return to insulation after each discharge. Possible causes of the fault include: (1) tree branches contacting the line, leading to intermittent discharge; (2) bird nest interference, such as intermittent contact between bird nest material and charged parts. In the preferred embodiment of this invention, "tree-line conflict" or "bird nest interference" is output as the fault cause identification result.

[0156] S5.3 For continuous periodic waveforms, the discharge process is characterized by continuous stable discharge, possibly accompanied by occasional strong discharge. Possible causes of the fault include: (1) dry tree branches forming a stable high-resistance grounding path on the insulated equipment, resulting in a continuous weak current; (2) small animals (such as squirrels, snakes, etc.) being electrocuted, creating a continuous grounding path. In the preferred embodiment of this invention, "high-resistance grounding of vegetation" or "small animal electrocution" is output as the fault cause identification result.

[0157] S5.4 For the spiky waveform, the discharge process is characterized by continuous weak discharge with a small but frequently fluctuating current amplitude. Possible causes of the fault include: aging of the transformer elbow joint insulation, causing multiple minor leakage currents. In this embodiment of the invention, the preferred output is "transformer elbow joint aging leakage current" as the fault cause identification result.

[0158] S5.5 For other waveforms, the discharge process may exhibit complex or irregular characteristics. Possible causes of failure include: (1) superposition of multiple fault types, such as the simultaneous existence of vegetation disturbance and equipment failure; (2) transient external interference, such as lightning strikes or brief contact with large animals. In the embodiments of the present invention, it is preferred to output "composite fault" or "external transient interference" as the result of fault cause identification.

[0159] S6 periodically collects new samples and retrains them to adapt to new data distributions and potential fault types.

[0160] The embodiments of the present invention retrain the GRU recurrent neural network by periodically updating the transient waveform data samples of single-phase grounding faults in the distribution network. The aim is to enable the model to adapt to new zero-sequence current waveforms and fault characteristics that may appear in the distribution network, thereby maintaining the accuracy and effectiveness of the fault diagnosis system in identifying single-phase grounding waveforms.

[0161] Compared with existing technologies, the embodiments of the present invention propose a systematic waveform classification standard to address the problem of inconsistent sampling frequency and sampling cycle number of transient waveform recording data in distribution networks. By analyzing the discharge process corresponding to each waveform morphology, a correlation model between waveform morphology and fault cause is established, providing a scientific basis for waveform morphology analysis.

[0162] The embodiments of this invention also propose a hierarchical fault diagnosis method from waveform morphology identification to fault cause identification. Based on the identification of single-phase grounding waveform morphology, the method makes fault diagnosis decisions by combining the correlation between waveform morphology and fault cause, which effectively improves the accuracy and interpretability of single-phase grounding fault diagnosis in distribution networks.

[0163] The embodiments of the present invention also apply the GRU recurrent neural network to waveform morphology recognition, effectively leveraging the GRU recurrent neural network's ability to extract temporal features of samples, realizing automated waveform morphology recognition, and improving recognition efficiency and accuracy.

[0164] The invention will now be described in conjunction with another alternative embodiment.

[0165] Example 2

[0166] The diagnostic device for the cause of power distribution network faults provided in this embodiment includes multiple implementation units, each of which corresponds to a specific implementation step in Embodiment 1 above.

[0167] Figure 4 This is a schematic diagram of an optional diagnostic device for the causes of power distribution network faults according to an embodiment of the present invention, such as... Figure 4 As shown, the device may include: an acquisition unit 41, a preprocessing unit 42, and an input unit 43.

[0168] The acquisition unit 41 is used to acquire fault current sequence data of the target power distribution line, wherein the fault current sequence data contains N fault current sampling data, where N is a positive integer.

[0169] The preprocessing unit 42 is used to preprocess the fault current sequence data to obtain the target time series data to be input into the model. The data preprocessing operations include data alignment, data normalization and time series reconstruction.

[0170] Input unit 43 is used to input target time series data into target model and output fault diagnosis results. The target model is a pre-trained recurrent network model. The target model is used to analyze the abnormal discharge process and identify the waveform shape of abnormal current based on target time series data. The fault diagnosis results are determined and output based on the waveform identification results.

[0171] The aforementioned diagnostic device for the causes of power distribution network faults can first acquire fault current sequence data of the target power distribution line through the acquisition unit 41. The fault current sequence data contains N fault current sampling data, where N is a positive integer. Then, the fault current sequence data is preprocessed by the preprocessing unit 42 to obtain the target time series data to be input into the model. The data preprocessing operations include data alignment, data normalization, and time series reconstruction. Finally, the target time series data is input into the target model through the input unit 43, and the fault diagnosis result is output. The target model is a pre-trained recurrent network model. The target model is used to analyze the abnormal discharge process and identify the waveform shape of the abnormal current based on the target time series data. The fault diagnosis result is determined and output based on the waveform identification result.

[0172] In this embodiment of the invention, fault current sequence data of the target distribution line containing rich fault information is obtained as the original information source. Preprocessing operations such as data alignment and data normalization reduce the impact of data errors caused by noise, amplitude differences, and sampling frequency differences of different devices in the original data. The time-series data reconstruction step can convert long sequence data into one-dimensional multi-channel time-series data as model input, which alleviates the gradient vanishing problem when recurrent neural networks process long sequence data, enabling the model to effectively learn long-term dependencies in the data. On this basis, the model can be used to extract multi-dimensional time-series features from the fault current sequence data, analyze the abnormal discharge process, and identify the waveform morphology of the abnormal current. Finally, the diagnostic result of the fault cause is output based on the waveform morphology. The above process makes full use of the ability of recurrent neural networks to process sequence data. By establishing a correlation between waveform morphology and fault cause, the fault type can be diagnosed more accurately based on the characteristics of the discharge process. This provides a more reliable and intelligent single-phase grounding fault diagnosis scheme for distribution networks, effectively avoiding the misdiagnosis problem caused by relying on only a single feature or simple threshold mapping, improving the overall reliability of the fault diagnosis system, and solving the technical problem that the fault cause diagnosis methods in related technologies often ignore the analysis of the discharge process and waveform, resulting in low reliability of the fault diagnosis results.

[0173] Optionally, the preprocessing unit includes: a first determining module, used to acquire a preset target frequency reduction frequency and a sampling frequency of the fault current sequence data, and determine a frequency reduction ratio based on the target frequency reduction frequency and the sampling frequency; and a data alignment module, used to perform data alignment processing on the fault current sequence data based on the frequency reduction ratio to obtain a frequency reduction sampling sequence, wherein the frequency reduction sampling sequence includes: M frequency reduction sampling data determined from N fault current sampling data, where M is a positive integer less than or equal to N.

[0174] Optionally, the preprocessing unit further includes: a second determining module, used to determine the maximum and minimum sampled data in the down-frequency sampling sequence; a first calculation module, used to calculate the normalized value of each down-frequency sampled data in the down-frequency sampling sequence based on the maximum and minimum sampled data, to obtain M normalized values; and a first arranging module, used to arrange the M normalized values ​​according to the temporal order in the down-frequency sampling sequence to obtain a normalized sequence.

[0175] Optionally, the preprocessing unit further includes: a first acquisition module, used to acquire the total number Q of sampling cycles in the fault current sequence data, where Q is a positive integer; a data segmentation module, used to segment the normalized sequence into Q segmented subsequences based on the sampling cycles, where each segmented subsequence contains P sequence data, where P is a positive integer; a third determination module, used to determine the feature data of each sampling cycle based on the segmented subsequences, where each sampling cycle corresponds to a time step of the target model; and a second arrangement module, used to arrange the feature data corresponding to the Q sampling cycles in temporal order to obtain the target temporal data after temporal reconstruction.

[0176] Optionally, the third determining module includes: an extraction submodule, used to extract P sequence data from the segmented subsequence for each sampling cycle; a linear superposition submodule, used to linearly superimpose the P sequence data to obtain the zero-sequence current time-series data corresponding to the sampling cycle; and a determining submodule, used to determine the zero-sequence current time-series data as the feature data of the sampling cycle.

[0177] Optionally, the target model consists of the following model layers: a batch normalization layer, used to receive preprocessed target time-series data and standardize the target time-series data; a recurrent layer, including a first gated recurrent unit, a first random deactivation unit, a second gated recurrent unit, and a second random deactivation unit, used to acquire standardized target time-series data and extract multidimensional time-series features from the target time-series data; and an output layer, including a batch normalization unit, a fully connected unit, and an activation function unit, used to map multidimensional time-series features to waveform morphology identification space to obtain waveform identification results, and determine and output fault diagnosis results based on the waveform identification results.

[0178] Optionally, the device for diagnosing the causes of power distribution network faults further includes: a second acquisition module, used to acquire a sample dataset and a pre-built initial model, wherein the initial model uses initial model parameters, and the sample dataset records transient waveform current data of S single-phase grounding samples in the power distribution system, where S is a positive integer; an iterative update module, used to divide the sample dataset according to a preset allocation ratio to obtain a training sample set and a test sample set, and use the training sample set to perform T rounds of iterative updates on the initial model, where T is a first preset value; a second calculation module, used to input the sample data from the training sample set into the model in R batches in each iteration, calculate the cross-entropy loss of the current batch and the gradient of the cross-entropy loss of the current batch with respect to each model parameter after each batch is input, and determine the model parameter update scheme for the current iteration based on the gradients of all batches, and perform model updates based on the model parameter update scheme, where R is a second preset value; and a fourth determination module, used to evaluate the accuracy of the latest model using the test sample set after performing T rounds of iterative updates, and determine the latest model as the target model if the model accuracy meets the standard.

[0179] The aforementioned diagnostic device for the causes of power distribution network faults may also include a processor and a memory. The aforementioned acquisition unit 41, preprocessing unit 42, input unit 43, etc., are all stored in the memory as program units, and the processor executes the aforementioned program units stored in the memory to realize the corresponding functions.

[0180] The aforementioned processor contains a kernel, which retrieves the corresponding program units from memory. One or more kernels can be configured. By adjusting kernel parameters, fault current sequence data of the target power distribution line can be obtained. This fault current sequence data contains N fault current sampling data points, where N is a positive integer. Data preprocessing is performed on the fault current sequence data to obtain the target time-series data to be input into the model. This target time-series data is then input into the target model, and the fault diagnosis result is output.

[0181] The aforementioned memory may include non-permanent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0182] This application also provides a computer program product, which, when executed on a data processing device, is suitable for executing an initialization program with the following method steps: acquiring fault current sequence data of a target power distribution line, wherein the fault current sequence data contains N fault current sampling data, where N is a positive integer; performing data preprocessing on the fault current sequence data to obtain target time-series data to be input into the model, wherein the data preprocessing operations include: data alignment, data normalization, and time-series reconstruction; inputting the target time-series data into the target model and outputting fault diagnosis results, wherein the target model is a pre-trained recurrent network model, the target model is used to analyze the abnormal discharge process and identify the waveform shape of the abnormal current based on the target time-series data, and determine and output the fault diagnosis results based on the waveform identification results.

[0183] According to another aspect of the present invention, a computer-readable storage medium is also provided, the computer-readable storage medium including a stored computer program, wherein, when the computer program is running, it controls the device where the computer-readable storage medium is located to execute the method for diagnosing the cause of a power distribution network fault as described in any of the above embodiments.

[0184] According to another aspect of the present invention, an electronic device is also provided, including one or more processors and a memory, wherein the memory is used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method for diagnosing the cause of power distribution network faults as described in any of the embodiments of the first invention.

[0185] Figure 5 This is a hardware structure block diagram of an electronic device (or mobile device) for a method of diagnosing the causes of faults in a power distribution network, according to an embodiment of the present invention. Figure 5 As shown, an electronic device may include one or more processors ( Figure 5 The image uses processors 502a, 502b, ..., 502n to illustrate the process. The processors may include, but are not limited to, processing devices such as microprocessors (MCUs) or programmable logic devices (FPGAs), and memory 504 for storing data. In addition, it may include: a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports in the I / O interface), a network interface, a keyboard, a power supply, and / or a camera. Those skilled in the art will understand that... Figure 5 The structure shown is for illustrative purposes only and does not limit the structure of the electronic device described above. For example, the electronic device may also include components that are more... Figure 5 The more or fewer components shown, or having the same Figure 5 The different configurations shown.

[0186] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0187] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0188] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0189] 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 units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0190] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0191] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, 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 storage medium 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 the present invention. The aforementioned storage medium 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.

[0192] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for diagnosing the causes of faults in a power distribution network, characterized in that, include: Obtain fault current sequence data of the target power distribution line, wherein the fault current sequence data contains N fault current sampling data, where N is a positive integer; The fault current sequence data is preprocessed to obtain the target time series data to be input into the model. The data preprocessing operations include: data alignment, data normalization, and time series reconstruction. The target time series data is input into the target model, and the fault diagnosis result is output. The target model is a pre-trained recurrent network model. The target model is used to analyze the abnormal discharge process and identify the waveform shape of the abnormal current based on the target time series data. The fault diagnosis result is determined and output based on the waveform identification result.

2. The diagnostic method according to claim 1, characterized in that, The steps for preprocessing the fault current sequence data include: Obtain the preset target frequency reduction frequency and the sampling frequency of the fault current sequence data, and determine the frequency reduction ratio based on the target frequency reduction frequency and the sampling frequency; Based on the down-frequency ratio, the fault current sequence data is aligned to obtain a down-frequency sampling sequence, wherein the down-frequency sampling sequence includes M down-frequency sampling data determined from the N fault current sampling data, where M is a positive integer less than or equal to N.

3. The diagnostic method according to claim 2, characterized in that, The steps for preprocessing the fault current sequence data include: Determine the maximum and minimum sampled data in the down-frequency sampling sequence; For each down-frequency sampled data in the down-frequency sampling sequence, a normalized value for the down-frequency sampled data is calculated based on the maximum sampled data and the minimum sampled data to obtain M normalized values; The M normalized values ​​are arranged according to the temporal order in the down-sampling sequence to obtain a normalized sequence.

4. The diagnostic method according to claim 3, characterized in that, The steps for preprocessing the fault current sequence data include: Obtain the total number Q of sampling cycles in the fault current sequence data, where Q is a positive integer; The normalized sequence is segmented into Q segmented subsequences based on the sampling cycle, wherein each segmented subsequence contains P sequence data, where P is a positive integer; For each of the sampling cycles, feature data of the sampling cycle are determined based on the segmented subsequence, wherein each of the sampling cycles corresponds to a time step of the target model; The feature data corresponding to the Q sampling cycles are arranged according to the time sequence to obtain the target time sequence data after time sequence reconstruction.

5. The diagnostic method according to claim 4, characterized in that, For each of the sampled cycles, the step of determining the characteristic data of that sampled cycle based on the segmented subsequence includes: For each sampling cycle, extract P sequence data from the segmented subsequence; The P sequence data are linearly superimposed to obtain the zero-sequence current time-series data corresponding to the sampling cycle; The zero-sequence current timing data is determined as the characteristic data of the sampling cycle.

6. The diagnostic method according to claim 1, characterized in that, The target model is composed of the following model layers: A batch normalization layer is used to receive the preprocessed target time series data and perform normalization processing on the target time series data. The recurrent layer includes a first gated recurrent unit, a first random deactivation unit, a second gated recurrent unit, and a second random deactivation unit, used to acquire the standardized target time series data and extract multidimensional time series features from the target time series data; The output layer includes a batch normalization unit, a fully connected unit, and an activation function unit, which are used to map the multidimensional time-series features to the waveform morphology identification space to obtain the waveform identification result, and determine and output the fault diagnosis result based on the waveform identification result.

7. The diagnostic method according to claim 1, characterized in that, The target model is obtained through the following steps: Obtain a sample dataset and a pre-built initial model, wherein the initial model uses initial model parameters, and the sample dataset records transient current data of S single-phase grounding samples in the power distribution system, where S is a positive integer; The sample dataset is divided according to a preset allocation ratio to obtain a training sample set and a test sample set. The initial model is then updated using the training sample set for T rounds, where T is a first preset value. In each iteration, the sample data from the training sample set is input into the model in R batches. After each batch is input, the cross-entropy loss of the current batch and the gradient of the cross-entropy loss of the current batch with respect to each model parameter are calculated. Based on the gradients of all batches, the model parameter update scheme for the current iteration is determined, and the model update is performed based on the model parameter update scheme. Here, R is a second preset value. After performing T rounds of iterative updates, the latest model is evaluated for accuracy using the test sample set. If the model accuracy meets the target, the latest model is determined as the target model.

8. A diagnostic device for the causes of power distribution network faults, characterized in that, include: The acquisition unit is used to acquire fault current sequence data of the target power distribution line, wherein the fault current sequence data includes N fault current sampling data, where N is a positive integer; The preprocessing unit is used to preprocess the fault current sequence data to obtain the target time series data to be input into the model. The data preprocessing operations include: data alignment, data normalization and time series reconstruction. The input unit is used to input the target time series data into the target model and output the fault diagnosis result. The target model is a pre-trained recurrent network model. The target model is used to analyze the abnormal discharge process and identify the waveform shape of the abnormal current based on the target time series data, and to determine and output the fault diagnosis result based on the waveform identification result.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device containing the computer-readable storage medium to perform the method for diagnosing the cause of a power distribution network fault as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, It includes one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method for diagnosing the causes of power distribution network faults as described in any one of claims 1 to 7.

Images