Distribution line fault type identification method based on data enhancement algorithm

Abstract

The invention relates to the technical field of power distribution network protection control, and discloses a power distribution line fault type identification method based on a data enhancement algorithm, and the method comprises the steps: obtaining fault time sequence data corresponding to each fault type of a power distribution line, and carrying out the processing, and obtaining three-phase voltage and three-phase current transient fault data of a long time scale; performing data enhancement on the long-time-scale three-phase voltage and three-phase current transient fault data to obtain a training data set; constructing a first power distribution network fault identification model; inputting the training data set into a first power distribution network fault recognition model for pre-training to obtain a second power distribution network fault recognition model; and based on the second power distribution network fault identification model, identifying the to-be-identified fault data of the power distribution line, and outputting the fault type of the power distribution line. The method has the effects of quickly judging the specific reason of the distribution line fault in the fault repair link of the power distribution network and meeting the requirement of high power supply reliability.

Description

Technical Field

[0001] This application relates to the field of power distribution network protection and control technology, and in particular to a method for identifying power distribution line fault types based on data augmentation algorithms. Background Technology

[0002] Currently, intelligent power distribution acquisition terminals and power distribution automation master stations have not only improved the operating efficiency and management level of the power grid, but have also basically realized fault location, fault isolation and power restoration of non-faulty areas in the power distribution network.

[0003] Analysis of the State Grid's construction and renovation projects reveals that high power supply reliability is a core indicator for future distribution networks. However, existing fault location systems can only pinpoint and isolate short-circuit and grounding faults in distribution lines, resulting in low efficiency in emergency repairs. Current technology cannot determine whether a line fault is caused by lightning strikes, insulator flashover, tree contact, or other types of faults, hindering the efficiency of on-site repair teams. Insufficient preparation of on-site repair measures or tools frequently leads to prolonged repair times. Furthermore, the lack of quantitative data support for distribution line reinforcement projects due to the long-standing absence of fault source diagnosis results in wasted investment or incomplete upgrades. While future distribution networks will offer flexible operation and strong resilient power supply capabilities, they are prone to power outages under extreme weather conditions. Once a power distribution line fails, in addition to locating the fault, it is also necessary to identify the fault and quickly determine the specific cause of the fault, such as lightning strikes causing line breakage, insulator flashover due to pollution, or trees touching the wire. This is so that a fault repair plan can be quickly formulated and repair materials can be allocated to guide subsequent power distribution line reinforcement plans, enabling rapid restoration of power supply loads and meeting the requirements for high power supply reliability.

[0004] Regarding the aforementioned technologies, the inventors have discovered that existing power distribution network fault repair processes suffer from the inability to determine the specific cause of power line faults, making it difficult to meet the requirements for high power supply reliability. Summary of the Invention

[0005] To quickly determine the specific cause of power distribution line faults, this application provides a power distribution line fault type identification method based on data augmentation algorithms.

[0006] Firstly, this application provides a method for identifying fault types in power distribution lines based on data augmentation algorithms.

[0007] This application is achieved through the following technical solution: A method for identifying fault types in power distribution lines based on data augmentation algorithms includes the following steps: Obtain fault time-series data corresponding to each fault type in the power distribution line, and process it to obtain long-term three-phase voltage and three-phase current transient fault data; Data augmentation was performed on long-term three-phase voltage and three-phase current transient fault data to obtain a training dataset; Constructing a first distribution network fault identification model, including: An input layer containing convolutional kernels of different sizes is used for parallel processing of fault data; A residual layer comprising several cascaded residual modules, the residual layer being used to reduce the sequence length of the input layer output; A channel attention mechanism for identifying important procedures in characteristic channels and a time attention mechanism for locating the time period of fault occurrence are introduced into the residual layer; An adaptive average pooling layer is used to pool the input sequence into the residual layer to a preset dimension, retaining global statistical information and providing a unified input dimension for the classifier. The second convolutional layer in the residual layer is replaced by the SKN model, focusing on the characteristic scales of three-phase voltage and three-phase current related to the fault type, to obtain the first distribution network fault identification model. The training dataset is input into the first distribution network fault identification model for pre-training to obtain the second distribution network fault identification model; Based on the second power distribution network fault identification model, the fault data to be identified in the power distribution line is identified, and the fault type of the power distribution line is output.

[0008] In a preferred embodiment, this application can be further configured to include the following steps: The fault data of the current distribution line is obtained for model migration and fine-tuning of the second distribution network fault identification model to obtain the target distribution network fault identification model; Replace the second distribution network fault identification model with the target distribution network fault identification model.

[0009] In a preferred embodiment, this application can be further configured as follows: the step of obtaining fault data of the current distribution line for model migration and fine-tuning of the second distribution network fault identification model to obtain the target distribution network fault identification model includes, Standardize the existing fault data in the distribution network to obtain real fault data; Freeze the weight parameters of the second distribution network fault identification model; Unfreeze the SKN module parameters of the second distribution network fault identification model, and replace the classifier module of the second distribution network fault identification model with a new, randomly initialized classifier module with an output dimension of 1. The real fault data is input into the second distribution network fault identification model for training, and the fine-tuning of the SKN module and the training of the classifier module are completed to obtain the target distribution network fault identification model.

[0010] In a preferred embodiment, this application may be further configured such that the channel attention mechanism includes a plurality of cascaded SE modules, wherein the final SE module is connected to the output of the last convolutional layer of the residual layer.

[0011] In a preferred embodiment, this application can be further configured such that: the time attention mechanism, based on the feature representation of the current time step, queries the feature representations of other time steps to obtain the degree of correlation between the feature representations of other time steps and the feature representation of the current time step; The feature representation of the current time step and the feature representation of other time steps are used as index keys for the query, and matched with the preset query vector to obtain the matching value. The feature representation of the time step corresponding to the largest matching value is taken as the critical fault occurrence period in the query vector.

[0012] In a preferred embodiment, this application can be further configured such that the step of performing data augmentation on the long-term three-phase voltage and three-phase current transient fault data to obtain a training dataset includes, Twenty-four TimeGAN models were initialized. The first TimeGAN model was used to enhance the A-phase voltage sequence fault data collected for a lightning-induced line breakage fault (without grounding). The second TimeGAN model was used to enhance the B-phase voltage sequence fault data collected for the same fault type. The third TimeGAN model was used to enhance the C-phase voltage sequence fault data collected for the same fault type. The fourth TimeGAN model was used to enhance the A-phase current sequence fault data collected for the same fault type. The fifth TimeGAN model was used to enhance the current sequence fault data collected for the same fault type. The first TimeGAN model enhances the B-phase current sequence fault data. The sixth TimeGAN model enhances the C-phase current sequence fault data collected for the line lightning-induced open-circuit fault type. The seventh TimeGAN model enhances the A-phase voltage sequence fault data collected for the line lightning-induced open-circuit fault type. The eighth TimeGAN model enhances the B-phase voltage sequence fault data collected for the line lightning-induced open-circuit fault type. The ninth TimeGAN model enhances the C-phase voltage sequence fault data collected for the line lightning-induced open-circuit fault type. The tenth TimeGAN model enhances the A-phase current sequence fault data collected for the line lightning-induced open-circuit fault type. The sequence fault data undergoes enhancement processing. The 11th TimeGAN model enhances the B-phase current sequence fault data obtained from the line lightning-induced open-circuit-to-ground fault type. The 12th TimeGAN model enhances the C-phase current sequence fault data obtained from the line lightning-induced open-circuit-to-ground fault type. The 13th TimeGAN model enhances the A-phase voltage sequence fault data obtained from the line insulator flashover fault type. The 14th TimeGAN model enhances the B-phase voltage sequence fault data obtained from the line insulator flashover fault type. The 15th TimeGAN model enhances the C-phase voltage sequence fault data obtained from the line insulator flashover fault type. The data undergoes augmentation processing. The 16th TimeGAN model enhances the A-phase current sequence fault data obtained from line insulator flashover faults; the 17th TimeGAN model enhances the B-phase current sequence fault data obtained from line insulator flashover faults; the 18th TimeGAN model enhances the C-phase current sequence fault data obtained from line insulator flashover faults; the 19th TimeGAN model enhances the A-phase voltage sequence fault data obtained from line tree contact faults; and the 20th TimeGAN model enhances the B-phase voltage sequence fault data obtained from line tree contact faults.The 21st TimeGAN model is used to enhance the C-phase voltage sequence fault data obtained from the tree contact wire fault type. The 22nd TimeGAN model is used to enhance the A-phase current sequence fault data obtained from the tree contact wire fault type. The 23rd TimeGAN model is used to enhance the B-phase current sequence fault data obtained from the tree contact wire fault type. The 24th TimeGAN model is used to enhance the C-phase current sequence fault data obtained from the tree contact wire fault type. Preset conditional judgment function and loss function, input the three-phase voltage and three-phase current transient fault data of the long time scale into 24 TimeGAN models for training, minimize the value of the loss function, and obtain 24 TimeGAN models for data augmentation of voltage sequence or current sequence of corresponding phase under each fault type of the distribution line. The training dataset was obtained by inputting long-term three-phase voltage and three-phase current transient fault data into 24 TimeGAN models for data augmentation.

[0013] In a preferred embodiment, this application can be further configured such that the step of inputting long-term three-phase voltage and three-phase current transient fault data into 24 TimeGAN models for data augmentation includes, Input the voltage or current sequence corresponding to the i-th TimeGAN model into the embedded component of the corresponding TimeGAN model, and use the conditional judgment function Sigmoid function to retain the feature information of the fault waveform of the distribution line to obtain the hidden state sequence. The hidden state sequence is input into the recovery component of the corresponding TimeGAN model to recover the feature information of the fault waveform of the power distribution line, and the fitted voltage sequence or current sequence is reconstructed and the fitted sequence is output. Minimize the value of the corresponding voltage or current sequence compared to the fitted sequence.

[0014] In a preferred embodiment, this application can be further configured such that the step of inputting long-term three-phase voltage and three-phase current transient fault data into 24 TimeGAN models for data augmentation includes, The hidden state sequence is input into the generator in the corresponding TimeGAN model to simulate the real fluctuation characteristics of the fault waveform data of the power distribution line and obtain the generated data sequence. Minimize the error between the generated data sequence and the hidden state sequence.

[0015] In a preferred embodiment, this application can be further configured such that the step of inputting long-term three-phase voltage and three-phase current transient fault data into 24 TimeGAN models for data augmentation includes, The hidden state sequence and the generated data sequence are input into the discriminator in the corresponding TimeGAN model; The discriminator is trained using a binary cross-entropy loss function.

[0016] In a preferred embodiment, this application can be further configured such that the step of inputting long-term three-phase voltage and three-phase current transient fault data into 24 TimeGAN models for data augmentation includes, The embedding component, the recovery component, the generator, and the discriminator are jointly optimized using a loss function that calculates the sum of the Euclidean distance, mean, and variance between the real data sequence and the generated data sequence until the sum of the Euclidean distance, mean, and variance between the real data sequence and the generated data sequence is minimized.

[0017] In summary, compared with the prior art, the beneficial effects of the technical solution provided in this application include at least the following: The system acquires and standardizes fault time-series data corresponding to various fault types in distribution lines to obtain long-term three-phase voltage and three-phase current transient fault data. Data augmentation is performed on the acquired fault time-series data for each fault type to address the issues of insufficient sample size and imbalanced sample size. A first distribution network fault identification model is constructed and pre-trained to identify the fault data to be identified in distribution lines and output the fault type of the distribution line. This solves the problem of insufficient fault sample data in the distribution network to be identified. Furthermore, the system utilizes the SKN model and attention mechanism to enhance the multi-scale feature adaptive extraction capability of the deep model under different types of faults in the distribution network. It focuses on the most discriminative feature scale in the distribution network to be identified, quickly determining the specific cause of the distribution line fault without manual fault location and identification. This enables rapid determination of the specific cause of the distribution line fault during the distribution network fault repair process, meeting the requirements for high power supply reliability. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall process of a power distribution line fault type identification method based on a data augmentation algorithm, provided as an exemplary embodiment of this application. Detailed Implementation

[0019] This specific embodiment is merely an explanation of this application and is not intended to limit it. After reading this specification, those skilled in the art can make modifications to this embodiment without contributing any inventive step, but such modifications are protected by patent law as long as they fall within the scope of the claims of this application.

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

[0021] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article, unless otherwise specified, generally indicates that the preceding and following related objects have an "or" relationship.

[0022] Existing distribution automation technologies can only locate, isolate, and restore power supply to non-faulty areas of the distribution network. Some new technologies can identify shallow faults such as single-phase grounding faults, two-phase short-circuit faults, two-phase grounding faults, and three-phase short-circuit faults. However, they cannot identify deep fault types such as line breaks caused by lightning strikes, line breaks caused by lightning strikes, insulator flashovers, or trees touching the ground. Therefore, they cannot effectively formulate emergency repair plans or scientifically and rationally reinforce the lines.

[0023] Therefore, this application provides a method for identifying fault types in power distribution lines based on data augmentation algorithms. The main steps of the method are described below.

[0024] Obtain fault time-series data corresponding to each fault type in the power distribution line, and process it to obtain long-term three-phase voltage and three-phase current transient fault data; Data augmentation was performed on long-term three-phase voltage and three-phase current transient fault data to obtain a training dataset; Constructing a first distribution network fault identification model, including: An input layer containing convolutional kernels of different sizes is used for parallel processing of fault data; A residual layer comprising several cascaded residual modules, the residual layer being used to reduce the sequence length of the input layer output; A channel attention mechanism for identifying important procedures in characteristic channels and a time attention mechanism for locating the time period of fault occurrence are introduced into the residual layer; An adaptive average pooling layer is used to pool the input sequence into the residual layer to a preset dimension, retaining global statistical information and providing a unified input dimension for the classifier. The second convolutional layer in the residual layer is replaced by the SKN model, focusing on the characteristic scales of three-phase voltage and three-phase current related to the fault type, to obtain the first distribution network fault identification model. The training dataset is input into the first distribution network fault identification model for pre-training to obtain the second distribution network fault identification model; Based on the second power distribution network fault identification model, the fault data to be identified in the power distribution line is identified, and the fault type of the power distribution line is output.

[0025] By using data augmentation algorithms and generative modeling techniques, the augmented massive amounts of multi-scenario data are input into the deep learning model for fault type identification. This constructs a data augmentation engine that can reflect the location of the power distribution line, the initial phase angle of each fault, and the fault type. This enables rapid and accurate identification of the time-domain waveform of power distribution line faults caused by various fault reasons, thereby improving the accuracy of power distribution line fault identification.

[0026] The embodiments of this application will now be described in further detail with reference to the accompanying drawings.

[0027] Reference Figure 1 Specifically, waveform data of single-phase grounding faults in distribution lines with available waveform recording data and confirmed fault types by the line inspection results are obtained. The fault types mainly include four types: line broken by lightning without falling to ground, line broken by lightning and falling to ground, insulator flashover, and tree contact with the wire, which are labeled as G1, G2, G3, and G4 respectively.

[0028] The acquired waveform data of single-phase ground faults in power distribution lines are standardized to reduce the difficulty of subsequent data processing. For data tagged G... i The fault type, obtain a certain first k Before the time of the secondary fault t 1. Duration and time after failure t The total duration is 2 hours. t 1+ t 2. Phase A voltage sequence U at the busbar at the feeder outlet i,k,A Phase B voltage sequence U i,k,B C-phase voltage sequence U i,k,C and the sequence of A-phase currents of each feeder I i,k,A The sum sequence of phase B currents I i,k,B C-phase current sum sequence I i,k,C To meet the needs of subsequent fault identification, unlike the time window requirements of traditional waveform data, the sampling interval Δ of these voltage and current fault data sequences is different. t It should be less than 100 microseconds. t 1. Not less than 10 seconds t 2. Not less than 10 seconds.

[0029] Data augmentation algorithms are employed to enhance the acquired fault time-series data for each fault type, addressing the issues of insufficient sample size and imbalanced sample size. For example, generative adversarial networks, random forests, or support vector machines can be used to generate fault waveform data of the same fault type under different fault locations, initial phase angles, and transition resistances, based on waveforms collected from real-world scenarios, thus supplementing the power distribution line fault waveform dataset for each fault type.

[0030] A fault identification model for the first distribution network is constructed and pre-trained by integrating Selective Kernel Networks (SKN) and Deep Residual Learning Networks (ResNet) models, including: Design a 1D ResNet model such that the fault type label obtained from data augmentation is G. i The three-phase voltage and three-phase current, a total of six data sequences, are used as input, and the fault type labels G are added. i As output.

[0031] To highlight the correlation between fault characteristics and fault types in each data sequence, an input module with six input channels was first designed, using convolutional kernels of different sizes, such as 7×7 or 3×3, to process the input signals in parallel. Then, four residual modules were constructed, connected in a cascaded manner (stage1-stage2-stage3-stage4). Each residual module contains multiple residual blocks, which are sequentially connected in series. The output of each block serves as the input of the next block. The first residual block of each residual module has a stride of 2 for downsampling, while the other residual blocks have a stride of 1. The number of fault characteristic channels was increased sequentially from 64 to 128, 256, and 512, with progressive downsampling to reduce the sequence length of the six data sequences (three-phase voltage and three-phase current). The number of channels in each residual module was 64-128-256-512, respectively.

[0032] The network architecture of each residual block includes: 1. One-dimensional convolution + regularization + ReLU activation function, 2. One-dimensional convolution + regularization, 3. Residual connection, 4. ReLU activation function.

[0033] Next, residual connections are used to solve the gradient vanishing problem. Channel attention in the ResNet model is used to first perform global average pooling to compress the entire spatial dimension (height × width) of each channel into a single value. Then, two fully connected layers are used to learn the dependencies between channels. Finally, a recalibrated feature map is output to identify which feature channels are most important for fault identification.

[0034] The network architecture for channel attention mainly consists of cascaded Squeeze-and-Excitation (SE) modules. In a standard residual block, the SE module is placed after the last convolutional layer and the batch normalization layer in the main path of the residual block.

[0035] By leveraging the temporal attention mechanism in the ResNet model, based on the feature representation of the current time step, queries are sent to other time steps to understand their relevance to the current time step.

[0036] The network architecture of temporal attention is mainly composed of a query-key-value mechanism, where the feature representations of all time steps serve as an "index" for the query and are used to perform matching calculations with the query vector.

[0037] The actual content information of all time steps is used to generate the final weighted output after obtaining attention weights, in order to locate the key fault occurrence period in the time series.

[0038] Temporal attention, channel attention, and the residual layer are connected in series. Input features first pass through the temporal attention module, which enhances the representation of important time steps and suppresses the influence of unimportant time steps. Then, the temporally enhanced features are fed into the residual layer module, which further reweights the importance of channels along the channel dimension to highlight task-relevant feature channels.

[0039] Finally, adaptive average pooling is used to pool the sequence features of different lengths obtained from the processing of the six data sequences into a fixed size, retaining the most important global statistical information and providing a unified input dimension for the classifier used for fault identification.

[0040] Adaptive average pooling serves as a preprocessing step for channel attention. First, the input feature map is transformed into a uniform, fixed size using adaptive average pooling to eliminate the influence of input scale variations. Then, the standardized pooled features are fed into the channel attention module. This allows the module to learn channel importance weights at a uniform scale, improving the stability and accuracy of weight learning.

[0041] Next, the residual modules in the constructed 1D ResNet model are replaced with the SKN model to obtain the first distribution network fault identification model based on the SKN and ResNet fusion model. The SKN model can automatically learn the optimal receptive field, focusing on the feature scales of the six data sequences of three-phase voltage and three-phase current, which are most relevant to the fault type. The first convolutional layer and the last residual connection part in the ordinary residual module of the ResNet model remain unchanged. The number of input and output channels of the SKN model is consistent with the original convolutional layer in the ordinary residual module of the ResNet model. The SKN model only replaces the second convolutional layer in the ordinary residual module of the ResNet model.

[0042] The SKN model consists of three key components: a split-transform-fusion selective kernel unit, a multi-branch feature extraction component, and an attention-guided feature selection component. These components are connected in series. By creating six convolutional branches with different kernel sizes, they are connected to the output of the first convolutional layer in the ordinary residual module of the ResNet model. After feature fusion and attention calculation of six data sequences of three-phase voltage and three-phase current, which are most relevant to the fault type, and weighted fusion, the results are fed into the residual connection part of the ordinary residual module in the ResNet model.

[0043] The data of the four types of distribution network faults after data augmentation are input into the first distribution network fault identification model for training, thereby realizing the training and establishment of the deep model for distribution network fault identification and obtaining the second distribution network fault identification model.

[0044] Based on the second distribution network fault identification model, the fault data to be identified in the distribution lines is identified, and the fault type of the distribution lines is output.

[0045] In one embodiment, a method for identifying fault types in power distribution lines based on data augmentation algorithms further includes the following steps: The fault data of the current distribution line is obtained for model migration and fine-tuning of the second distribution network fault identification model to obtain the target distribution network fault identification model; Replace the second distribution network fault identification model with the target distribution network fault identification model.

[0046] Based on a small number of existing fault data samples in the distribution network, the second distribution network fault identification model performs model migration and fine-tuning, and outputs any one of the following fault types: line lightning strike disconnection without grounding, line lightning strike disconnection with grounding, insulator flashover, and tree contact with wire, thereby realizing the fault identification work of the fault to be identified in the distribution network.

[0047] Specifically, a small number of existing fault data samples in the distribution network to be identified are collected and the data samples are standardized.

[0048] Freeze the model weight parameters of the second distribution network fault identification model.

[0049] Unfreeze the SKN module parameters in the second distribution network fault identification model and replace the classifier in the second distribution network fault identification model with a new, randomly initialized classifier module with an output dimension of 1.

[0050] A small number of fault data samples from the current distribution network are input into the corrected second distribution network fault identification model for retraining. Based on the training process, the SKN module is fine-tuned and the new classifier is trained to obtain the target distribution network fault identification model.

[0051] Through model transfer and fine-tuning, the model can be retrained for samples in the distribution network to be identified, thereby improving the adaptability of the second distribution network fault identification model based on the SKN and ResNet fusion model and reducing overfitting.

[0052] When a fault occurs in a distribution line of the distribution network that needs to be identified, the fault data of the distribution line is acquired and the fault time sequence data corresponding to each fault type of the distribution line is standardized. Then, the six data sequences of three-phase voltage and three-phase current, which are most relevant to the fault type, are input into the second distribution network fault identification model. Based on the model output results, the fault type is determined to be one of the following: line broken by lightning without falling to the ground, line broken by lightning and falling to the ground, insulator flashover, or tree contact with the wire.

[0053] In one embodiment, to address the issues of small sample size and imbalanced sample size, data augmentation is performed on the acquired fault time-series data for each fault type based on the TimeGAN (Time-series Generative Adversarial Networks) model to obtain a training dataset, including: Initialize 24 TimeGAN models. The first TimeGAN model is used to obtain the A-phase voltage sequence U for line lightning strike-induced line breakage without grounding fault types. i,k,A The fault data is augmented, and the second TimeGAN model is used to analyze the B-phase voltage sequence U obtained from the lightning strike-induced line breakage without grounding fault type. i,k,B The fault data is augmented, and the third TimeGAN model is used to analyze the C-phase voltage sequence U obtained from the lightning strike-induced line breakage without grounding fault type. i,k,C The fault data is augmented, and the fourth TimeGAN model is used to analyze the A-phase current sequence I obtained from the lightning strike-induced line breakage without grounding fault type. i,k,A The fault data is augmented, and the fifth TimeGAN model is used to analyze the B-phase current sequence I obtained from the lightning strike-induced line breakage without grounding fault type. i,k,B The fault data is enhanced, and the sixth TimeGAN model is used to analyze the C-phase current sequence I obtained from the lightning strike-induced line breakage without grounding fault type. i,k,C The fault data is augmented; the 7th TimeGAN model is used to analyze the A-phase voltage sequence U obtained from the lightning strike ground fault type. i,k,A The fault data is augmented; the 8th TimeGAN model is used to analyze the B-phase voltage sequence U obtained from the lightning strike-induced line breakage and ground fault type. i,k,B The fault data is augmented, and the 9th TimeGAN model is used to analyze the C-phase voltage sequence U obtained from the lightning strike ground fault type. i,k,CThe fault data is augmented, and the 10th TimeGAN model is used to analyze the A-phase current sequence I obtained from the lightning strike-induced line breakage and ground fault type. i,k,A The fault data is augmented, and the 11th TimeGAN model is used to analyze the B-phase current sequence I obtained from the lightning strike ground fault type. i,k,B The fault data is augmented, and the 12th TimeGAN model is used to analyze the C-phase current sequence I obtained from the lightning strike-induced line breakage and ground fault type. i,k,C The fault data is enhanced; the 13th TimeGAN model is used to analyze the A-phase voltage sequence U obtained from the line insulator flashover fault type. i,k,A The fault data is enhanced, and the 14th TimeGAN model is used to analyze the B-phase voltage sequence U obtained from the line insulator flashover fault type. i,k,B The fault data is enhanced, and the 15th TimeGAN model is used to analyze the C-phase voltage sequence U obtained from the line insulator flashover fault type. i,k,C The fault data is enhanced, and the 16th TimeGAN model is used to analyze the A-phase current sequence I obtained from the line insulator flashover fault type. i,k,A The fault data is enhanced, and the 17th TimeGAN model is used to analyze the B-phase current sequence I obtained from the line insulator flashover fault type. i,k,B The fault data is enhanced, and the 18th TimeGAN model is used to analyze the C-phase current sequence I obtained from the line insulator flashover fault type. i,k,C The fault data is augmented; the 19th TimeGAN model is used to analyze the A-phase voltage sequence U obtained from the tree contact fault type on the line. i,k,A The fault data is augmented, and the 20th TimeGAN model is used to analyze the B-phase voltage sequence U obtained from the tree contact fault type on the line. i,k,B The fault data is augmented, and the 21st TimeGAN model is used to analyze the C-phase voltage sequence U obtained from the tree contact fault type on the line. i,k,C The fault data is augmented, and the 22nd TimeGAN model is used to analyze the A-phase current sequence I obtained from the tree contact fault type on the line. i,k,A The fault data is augmented, and the 23rd TimeGAN model is used to analyze the B-phase current sequence I obtained from the tree contact fault type on the line. i,k,B The fault data is augmented, and the 24th TimeGAN model is used to analyze the C-phase current sequence I obtained from the tree contact fault type on the line. i,k,C Enhancement processing is performed on fault data; Preset conditional judgment function and loss function, input the three-phase voltage and three-phase current transient fault data of the long time scale into 24 TimeGAN models for training, minimize the value of the loss function, and obtain 24 TimeGAN models for data augmentation of voltage sequence or current sequence of corresponding phase under each fault type of the distribution line. The training dataset was obtained by inputting long-term three-phase voltage and three-phase current transient fault data into 24 TimeGAN models for data augmentation.

[0054] The steps of inputting the long-term three-phase voltage and three-phase current transient fault data into 24 TimeGAN models for training include: The voltage or current sequence corresponding to the i-th TimeGAN model is input into the embedded component of the model, and the conditional judgment function f1: Sigmoid function is used to retain the feature information of the distribution line fault waveform to obtain the hidden state sequence H. i ; The hidden state sequence H i The recovery component in the input model recovers the feature information of the fault waveform of the power distribution line and reconstructs the fitted voltage or current sequence. The model is trained with the goal of minimizing the value of the input true sequence and the reconstructed fitted sequence, thereby improving the ability of the embedded component and recovery component in the TimeGAN model to retain and recover the feature information of the fault waveform of the power distribution line. And the hidden state sequence H obtained from the i-th TimeGAN model i Input the generator of the i-th TimeGAN model, and minimize the difference between the generator-generated data sequence and the hidden state sequence H by leveraging the generator's predictive power for the next step. i The error between the two is used to train the generator, so as to improve the generator's ability to simulate the real fluctuation characteristics of power distribution line fault waveform data in the i-th TimeGAN model. Next, the hidden state sequence H obtained from the i-th TimeGAN model is... i The generated data sequence of the generator is simultaneously input into the discriminator of the i-th TimeGAN model. The discriminator is trained by the binary classification cross-entropy loss function to improve the discriminator's ability to identify the hidden state sequence of the real fault waveform in the i-th TimeGAN model. Furthermore, the embedded components, generator, and discriminator in the i-th TimeGAN model are jointly optimized. By using conditional judgment function f1 and loss function f2, the embedding component, generator, and discriminator are jointly trained. The gradients of the generator and discriminator are used to fine-tune the embedding component and recovery component, so that the sum of the Euclidean distance, mean, and variance of the generator's generated data sequence and the real fault sequence is minimized. This ensures that the hidden state sequence generated by the generator will not be identified as false data by the discriminator. That is, the conditional judgment function f1 needs to output "true" to achieve the enhancement of distribution network fault data.

[0055] The expressions for the conditional judgment function f1 and the loss function f2 are as follows: In the formula, These represent the real sequence and the generated sequence, respectively, with MSE being the Euclidean distance function. The global mean. For global variance, Hyperparameters, used to quantify the importance of the three factors, are set by professionals.

[0056] Based on the loss function, the TimeGAN model can adaptively demodulate the weights between neurons to minimize the error between the input true sequence and the reconstructed fitted sequence, and to minimize the error between the generator-generated data sequence and the hidden state sequence.

[0057] The 24 TimeGAN models were trained sequentially to obtain TimeGAN models that augmented the voltage or current sequences of each phase for each fault type in the power distribution line.

[0058] In this embodiment, 500 sets of data waveforms for each phase voltage and current are obtained under four fault types: line lightning strike without grounding, line lightning strike with grounding, insulator flashover, and tree contact. This ensures the balance of data samples for each fault type. Of course, the specific number of sets can be set according to actual needs.

[0059] The TimeGAN model can generate fault waveform data of the same fault type under different fault locations, different initial phase angles of faults, and different transition resistances based on waveforms collected from real-world fields. This supplements the power distribution line fault waveform dataset for this fault type and solves the problems of small sample size and imbalanced sample size.

[0060] In summary, a data augmentation-based method for identifying distribution line fault types involves acquiring and standardizing the fault time-series data corresponding to each fault type in the distribution line to obtain long-term three-phase voltage and three-phase current transient fault data. Then, based on the TimeGAN model, data augmentation is performed on the standardized fault time-series data of each fault type to address the problems of small sample size and imbalanced samples. Next, selective kernel network and deep residual network models are integrated to realize the construction and pre-training of the distribution network fault type identification model. Finally, based on the limited fault data samples already available in the distribution network, the distribution network fault type identification model is transferred and fine-tuned to achieve the identification of distribution network faults to be identified.

[0061] A data augmentation-based method for identifying distribution line fault types addresses the problem of insufficient fault sample data in the distribution network to be identified. It utilizes the SKN attention mechanism to enhance the multi-scale feature adaptive extraction capability of the deep model under different types of faults in the distribution network, focusing on the most discriminative feature scale in the distribution network to be identified. Furthermore, it employs transfer learning to retrain the model on samples from the distribution network to be identified, thereby improving the adaptability of the distribution network fault type identification model and reducing overfitting.

[0062] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0063] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable 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 this application. When executed, the computer program may include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0064] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above.

Claims

1. A method for identifying fault types in power distribution lines based on data augmentation algorithms, characterized in that, Includes the following steps, Obtain fault time-series data corresponding to each fault type in the power distribution line, and process it to obtain long-term three-phase voltage and three-phase current transient fault data; Data augmentation was performed on long-term three-phase voltage and three-phase current transient fault data to obtain a training dataset; Constructing a first distribution network fault identification model, including: An input layer containing convolutional kernels of different sizes is used for parallel processing of fault data; A residual layer comprising several cascaded residual modules, the residual layer being used to reduce the sequence length of the input layer output; A channel attention mechanism for identifying important procedures in characteristic channels and a time attention mechanism for locating the time period of fault occurrence are introduced into the residual layer; An adaptive average pooling layer is used to pool the input sequence into the residual layer to a preset dimension, retaining global statistical information and providing a unified input dimension for the classifier. The second convolutional layer in the residual layer is replaced by the SKN model, focusing on the characteristic scales of three-phase voltage and three-phase current related to the fault type, to obtain the first distribution network fault identification model. The training dataset is input into the first distribution network fault identification model for pre-training to obtain the second distribution network fault identification model; Based on the second power distribution network fault identification model, the fault data to be identified in the power distribution line is identified, and the fault type of the power distribution line is output.

2. The method for identifying power distribution line fault types based on data augmentation algorithms according to claim 1, characterized in that, It also includes the following steps, The fault data of the current distribution line is obtained for model migration and fine-tuning of the second distribution network fault identification model to obtain the target distribution network fault identification model; Replace the second distribution network fault identification model with the target distribution network fault identification model.

3. The method for identifying power distribution line fault types based on data augmentation algorithms according to claim 2, characterized in that, The step of obtaining fault data of the current distribution line for model migration and fine-tuning of the second distribution network fault identification model to obtain the target distribution network fault identification model includes the following steps: Standardize the existing fault data in the distribution network to obtain real fault data; Freeze the weight parameters of the second distribution network fault identification model; Unfreeze the SKN module parameters of the second distribution network fault identification model, and replace the classifier module of the second distribution network fault identification model with a new, randomly initialized classifier module with an output dimension of 1. The real fault data is input into the second distribution network fault identification model for training, and the fine-tuning of the SKN module and the training of the classifier module are completed to obtain the target distribution network fault identification model.

4. The method for identifying power distribution line fault types based on data augmentation algorithms according to claim 1, characterized in that, The channel attention mechanism includes several cascaded SE modules, wherein the last SE module is connected to the output of the last convolutional layer of the residual layer.

5. The method for identifying power distribution line fault types based on data augmentation algorithms according to claim 1, characterized in that, The time attention mechanism, based on the feature representation of the current time step, queries the feature representations of other time steps to obtain the degree of correlation between the feature representations of other time steps and the feature representation of the current time step. The feature representation of the current time step and the feature representation of other time steps are used as index keys for the query, and matched with the preset query vector to obtain the matching value. The feature representation of the time step corresponding to the largest matching value is taken as the critical fault occurrence period in the query vector.

6. The method for identifying distribution line fault types based on data augmentation algorithms according to any one of claims 1-5, characterized in that, The steps of performing data augmentation on long-term three-phase voltage and three-phase current transient fault data to obtain a training dataset include: Twenty-four TimeGAN models were initialized. The first TimeGAN model was used to enhance the A-phase voltage sequence fault data collected for a lightning-induced line breakage fault (without grounding). The second TimeGAN model was used to enhance the B-phase voltage sequence fault data collected for the same fault type. The third TimeGAN model was used to enhance the C-phase voltage sequence fault data collected for the same fault type. The fourth TimeGAN model was used to enhance the A-phase current sequence fault data collected for the same fault type. The fifth TimeGAN model was used to enhance the current sequence fault data collected for the same fault type. The first TimeGAN model enhances the B-phase current sequence fault data. The sixth TimeGAN model enhances the C-phase current sequence fault data collected for the line lightning-induced open-circuit fault type. The seventh TimeGAN model enhances the A-phase voltage sequence fault data collected for the line lightning-induced open-circuit fault type. The eighth TimeGAN model enhances the B-phase voltage sequence fault data collected for the line lightning-induced open-circuit fault type. The ninth TimeGAN model enhances the C-phase voltage sequence fault data collected for the line lightning-induced open-circuit fault type. The tenth TimeGAN model enhances the A-phase current sequence fault data collected for the line lightning-induced open-circuit fault type. The sequence fault data undergoes enhancement processing. The 11th TimeGAN model enhances the B-phase current sequence fault data obtained from the line lightning-induced open-circuit-to-ground fault type. The 12th TimeGAN model enhances the C-phase current sequence fault data obtained from the line lightning-induced open-circuit-to-ground fault type. The 13th TimeGAN model enhances the A-phase voltage sequence fault data obtained from the line insulator flashover fault type. The 14th TimeGAN model enhances the B-phase voltage sequence fault data obtained from the line insulator flashover fault type. The 15th TimeGAN model enhances the C-phase voltage sequence fault data obtained from the line insulator flashover fault type. The data undergoes augmentation processing. The 16th TimeGAN model enhances the A-phase current sequence fault data obtained from line insulator flashover faults; the 17th TimeGAN model enhances the B-phase current sequence fault data obtained from line insulator flashover faults; the 18th TimeGAN model enhances the C-phase current sequence fault data obtained from line insulator flashover faults; the 19th TimeGAN model enhances the A-phase voltage sequence fault data obtained from line tree contact faults; and the 20th TimeGAN model enhances the B-phase voltage sequence fault data obtained from line tree contact faults.The 21st TimeGAN model is used to enhance the C-phase voltage sequence fault data obtained from the tree contact wire fault type. The 22nd TimeGAN model is used to enhance the A-phase current sequence fault data obtained from the tree contact wire fault type. The 23rd TimeGAN model is used to enhance the B-phase current sequence fault data obtained from the tree contact wire fault type. The 24th TimeGAN model is used to enhance the C-phase current sequence fault data obtained from the tree contact wire fault type. Preset conditional judgment function and loss function, input the three-phase voltage and three-phase current transient fault data of the long time scale into 24 TimeGAN models for training, minimize the value of the loss function, and obtain 24 TimeGAN models for data augmentation of voltage sequence or current sequence of corresponding phase under each fault type of the distribution line. The training dataset was obtained by inputting long-term three-phase voltage and three-phase current transient fault data into 24 TimeGAN models for data augmentation.

7. The method for identifying distribution line fault types based on data augmentation algorithms according to claim 6, characterized in that, The steps of inputting long-term three-phase voltage and three-phase current transient fault data into 24 TimeGAN models for data augmentation include: Input the voltage or current sequence corresponding to the i-th TimeGAN model into the embedded component of the corresponding TimeGAN model, and use the conditional judgment function Sigmoid function to retain the feature information of the fault waveform of the distribution line to obtain the hidden state sequence. The hidden state sequence is input into the recovery component of the corresponding TimeGAN model to recover the feature information of the fault waveform of the power distribution line, and the fitted voltage sequence or current sequence is reconstructed and the fitted sequence is output. Minimize the value of the corresponding voltage or current sequence compared to the fitted sequence.

8. The method for identifying power distribution line fault types based on data augmentation algorithms according to claim 7, characterized in that, The steps of inputting long-term three-phase voltage and three-phase current transient fault data into 24 TimeGAN models for data augmentation include: The hidden state sequence is input into the generator in the corresponding TimeGAN model to simulate the real fluctuation characteristics of the fault waveform data of the power distribution line and obtain the generated data sequence. Minimize the error between the generated data sequence and the hidden state sequence.

9. The method for identifying power distribution line fault types based on data augmentation algorithms according to claim 8, characterized in that, The steps of inputting long-term three-phase voltage and three-phase current transient fault data into 24 TimeGAN models for data augmentation include: The hidden state sequence and the generated data sequence are input into the discriminator in the corresponding TimeGAN model; The discriminator is trained using a binary cross-entropy loss function.

10. The method for identifying distribution line fault types based on data augmentation algorithms according to claim 9, characterized in that, The steps of inputting long-term three-phase voltage and three-phase current transient fault data into 24 TimeGAN models for data augmentation include: The embedding component, the recovery component, the generator, and the discriminator are jointly optimized using a loss function that calculates the sum of the Euclidean distance, mean, and variance between the real data sequence and the generated data sequence until the sum of the Euclidean distance, mean, and variance between the real data sequence and the generated data sequence is minimized.

Images