A method for fault data identification in distribution network based on convolutional neural network

Abstract

The present invention relates to a method for identifying distribution network fault data based on convolutional neural network, and belongs to the field of electric power technology. The method comprises: S1: obtaining transient fault data of distribution network and preprocessing it; S2: using multi-layer 1DCAE to build a fault feature extraction network, and inputting the preprocessed transient fault data to train and optimize the fault feature extraction network to obtain low-dimensional fault features; S3: using multi-layer 1DCNN to build a fault feature classification model, and inputting low-dimensional fault features to train and optimize the fault feature classification model to complete the identification of different types of transient fault data. The present invention has automatic feature extraction capability and good fault tolerance capability, and is more suitable for the identification of massive fault data in distribution network.

Description

Technical Field

[0001] The invention belongs to the field of electric power technology and relates to a distribution network fault data identification method based on convolutional neural network. Background Art

[0002] As the power system moves towards intelligence and transparency, the structure of the distribution network is becoming increasingly complex, the detection equipment is becoming more numerous, and the probability of various faults occurring during operation is increasing sharply. After a fault occurs, in order to avoid further deterioration and spread of the fault, it is necessary to quickly and accurately locate the fault, reconfigure the network, merge the fault, analyze the fault event, and troubleshoot and repair it. Fault diagnosis and identification are the prerequisites for the above work.

[0003] At present, the fault diagnosis and classification methods for distribution networks can be generally classified into three categories:

[0004] The first category is the traditional fault identification method based on knowledge analysis, which has been used in the field of power system fault identification for a long time, but it still has problems such as weak fault tolerance, difficult operation, and low identification efficiency.

[0005] The second category is a composite fault classification and identification model that combines traditional fault diagnosis methods with a new generation of artificial intelligence. There are two common types of combinations: one is to use the classic fault signal processing method to extract features and then combine it with an artificial intelligence deep learning model for fault classification and identification; the other is to use the data feature self-learning ability of the neural network to extract fault features and then combine it with the traditional shallow classification network for fault diagnosis. This type of method improves the classification performance to a certain extent, but manual feature analysis and extraction takes a lot of time, the model is easily affected by human factors, and manual feature extraction faces greater difficulties for multi-type fault samples with small differences.

[0006] The third category is that both data feature processing and feature classifiers use a deep learning framework to identify fault types. This type of full deep learning model is more suitable for identifying massive fault data in distribution networks, but the research on this model is not yet complete and its adaptability is relatively weak.

[0007] Therefore, there is an urgent need for a distribution network fault data identification method that can solve the problems of low fault tolerance and low recognition efficiency of traditional artificial feature extraction methods, as well as low identification accuracy of SVM, ANN and CNN. Summary of the invention

[0008] In view of this, the purpose of the present invention is to provide a distribution network fault data identification method based on convolutional neural network, solve the difficulty of self-extraction of measured transient fault data features, and then input the extracted data features into a one-dimensional convolutional neural network to achieve the goal of end-to-end fault data identification intelligently and efficiently. Compared with the existing methods, it has automatic feature extraction capabilities and better fault tolerance, and is more suitable for massive fault data identification in distribution networks.

[0009] In order to achieve the above object, the present invention provides the following technical solutions:

[0010] A method for identifying distribution network fault data based on convolutional neural network specifically comprises the following steps:

[0011] S1: Obtain transient fault data of distribution network and pre-process it;

[0012] S2: Use a multi-layer one-dimensional convolutional auto-encoder (1DCAE) to build a fault feature extraction network, and input the pre-processed transient fault data to train and optimize the fault feature extraction network to obtain low-dimensional fault features;

[0013] S3: Use a multi-layer one-dimensional convolutional neural network (1DCNN) to build a fault feature classification model, and input low-dimensional fault features to train and optimize the fault feature classification model to complete the recognition of different types of transient fault data.

[0014] Further, step S1 specifically includes: obtaining the electrical quantity information of the fault line recorded in the massive wave recording file of the master station, that is, the transient fault data of the distribution network; the preprocessing stage is to intercept the electrical quantity information near the fault point that best reflects the fault characteristics as the training data sample of the network.

[0015] Further, in step S1, preprocessing specifically includes: the obtained distribution network transient fault data is composed of 256 three-phase voltage sampling data before and after the fault moment, so as to accurately obtain the three-phase voltage change characteristics at the fault moment; and perform certain data sample expansion such as translation and noise addition.

[0016] Furthermore, step S2 specifically includes: building a fault feature extraction network by stacking multiple layers of 1DCAE, and uniformly using the preprocessed transient fault data of different types for network encoding and decoding training, adjusting the network parameters to optimize the network weights and biases so that the network has the best feature extraction effect, and then outputting low-dimensional fault features from the encoder for fault data classification and identification in step S3.

[0017] Further, in step S2, the 1DCAE is composed of a convolutional layer, a pooling layer, a depooling layer and a deconvolutional layer; the convolutional pooling data is reconstructed by deconvolution and depooling using the spatial information preservation characteristics of convolution, and then the model is solved by optimizing the minimum error between the reconstructed data and the original data, and finally the encoded low-dimensional effective features are extracted.

[0018] Furthermore, step S3 specifically includes: relearning the low-dimensional fault features obtained in step S2 to train the optimal parameters of the network; then classifying the fault features through the convolutional network Softmax classifier layer to complete the recognition of different types of transient fault data.

[0019] Furthermore, in step S3, the 1DCNN is composed of a one-dimensional convolutional layer, a one-dimensional pooling layer, a fully connected layer and a classification layer, and is coordinated with an activation function, regularization, an optimizer, etc. to complete training model tuning.

[0020] Furthermore, in step S3, the fully connected layer in the 1DCNN is generated by converting the data block after convolution pooling into one dimension, and the dropout method is added in front of it to randomly delete some neurons to suppress model overfitting, and the multi-classification of data can be completed in conjunction with the classifier.

[0021] Further, in step S3, the output of the fully connected layer is:

[0022]

[0023] in, is the output of the fully connected layer, is the fully connected layer input, ω is the weight, is the bias, f(·) is the activation function, and l is the number of fully connected layers.

[0024] The classifier uses the Softmax classifier with a fully connected layer to complete data multi-classification, and the expression is:

[0025]

[0026] Among them, y j is the jth output neuron, and n is the number of neurons.

[0027] The beneficial effects of the present invention are as follows: the network architecture of the present invention uses artificial intelligence deep learning to automatically perform diagnostic analysis from fault data preprocessing to fault data feature extraction and then to fault feature classification, reducing interference from human factors and directly obtaining the mapping relationship between fault type and fault feature from the original recorded fault data, thereby avoiding the traditional manual analysis of fault data.

[0028] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0030] Figure 1 It is a schematic diagram of the 1DCAE structure;

[0031] Figure 2 Schematic diagram of 1DCNN structure;

[0032] Figure 3 This is a framework diagram for classification and identification of end-to-end transient fault data;

[0033] Figure 4 Visualize results for data features;

[0034] Figure 5 This is the comparative training result of the 1DCAE-1DCNN classification model of the present invention and the 1DCNN classification model. DETAILED DESCRIPTION

[0035] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0036] See also Figure 1 to Figure 5 , which is a distribution network fault data identification method based on convolutional neural network, mainly includes the following three parts:

[0037] 1) Fault data preprocessing

[0038] Since the transient waveform fault indicator will record the electrical quantity information of the 4 cycles before and after the fault point when recording the fault, and the data points around the fault point can best reflect the fault information, the sample data selected is composed of 256 three-phase voltage sampling data before and after the fault moment to accurately obtain the three-phase voltage change characteristics at the moment of the fault. Considering that the sample size of the selected fault type in the data set is insufficient, a certain amount of data sample expansion such as translation and noise addition is performed to form a total of 1,800 data sets of six types of faults. According to the sampling frequency of the fault indicator, the dimension of each network training sample data is 768.

[0039] 2) Multi-layer 1DCAE builds a fault feature extraction network

[0040] 1DCAE uses convolutional layers instead of ordinary fully connected layers. Its structure is as follows Figure 1 , which consists of a convolutional layer, a pooling layer, a depooling layer, and a deconvolutional layer. The idea is to use the spatial information preservation characteristics of convolution to reconstruct the convolutional pooled data through deconvolution and depooling, and then solve its model by optimizing the minimum error between the reconstructed data and the original data, and finally extract the encoded low-dimensional effective features. In the feature extraction process, its local perception and weight sharing characteristics can enable the model to obtain more accurate data features and have more efficient computing efficiency. The one-dimensional transient recording fault data of the distribution network has certain local spatial connections and higher dimensions, so 1DCAE is more suitable for the efficient extraction of transient fault data features.

[0041] 3) Fault feature classification model built with multi-layer 1DCNN

[0042] The convolutional neural network is inspired by the way mammalian visual receptive fields transmit retinal visual information. It is a deep learning model that imitates the excitation state of animals' multi-layer receptive fields by stacking multiple convolutional layers and pooling layers. Figure 2 It is a 1DCNN structure, consisting of a one-dimensional convolution layer, a one-dimensional pooling layer, a fully connected layer, and a classification layer, and is coordinated with activation functions, regularization, optimizers, etc. to complete the training model tuning. The local perception characteristics of the convolutional neural network enable each neuron to focus on responding to the feature information within its own receptive field, so that the convolution kernel has the best response to the spatial local pattern of the input data; its weight sharing characteristics enable different feature surfaces in the convolution layer to share a unified convolution kernel, optimize the data connection between the layers of the network, and reduce the complexity of the network structure and the number of weights.

[0043] Fully connected layer: After convolution and pooling, the data block is converted into a one-dimensional fully connected layer. Usually, the dropout method is added in front of it to randomly delete some neurons to suppress model overfitting. In combination with the classifier, multiple classification of data can be completed. The output of the fully connected layer is:

[0044]

[0045] in, is the output of the fully connected layer, is the fully connected layer input, ω is the weight, is the bias, f(·) is the activation function, and l is the number of fully connected layers.

[0046] Classifier: Use the Softmax classifier with the fully connected layer to complete data multi-classification. The expression is:

[0047]

[0048] Among them, y j is the jth output neuron, and n is the number of neurons.

[0049] Embodiment 1:

[0050] This embodiment provides an end-to-end transient fault data classification and recognition method based on a dual convolutional neural network. The principle framework of the method is as follows: Figure 3 As shown, the entire architecture mainly consists of three parts.

[0051] The first part is the preprocessing stage of the transient waveform recording fault data of the distribution network. The massive waveform recording files transmitted to the master station record the electrical quantity information of the faulty line. The preprocessing stage is to intercept the electrical quantity information near the fault point that best reflects the fault characteristics as the network training data sample.

[0052] The second part builds a fault feature extraction network by stacking multiple layers of 1DCAE. In the fault data set, different types of transient fault data are uniformly used for network encoding and decoding training. The network parameters are adjusted to optimize the network weights and biases so that the network has the best feature extraction effect. Then, the encoder outputs low-dimensional fault features for data classification and recognition in the third part.

[0053] The third part is a fault feature classification model built using multi-layer 1DCNN. The low-dimensional fault features of the original fault data are re-learned and the optimal parameters of the network are trained. Then the fault features are classified through the convolutional network Softmax classifier layer to complete the recognition of different types of transient fault data.

[0054] Simulation experiment:

[0055] 1) Fault data feature extraction

[0056] Taking the fault recording data of a substation as an example, 1DCAE is used to extract features from the fault data samples. The experiment uses Google's second-generation machine learning system TensorFlow, the development language uses Python, the computer configuration is Inter Corei7-10700CPU@2.9GHz, and the memory is 8GB. First, a one-dimensional convolutional autoencoder data flow graph is built in the TensorFlow framework to obtain the convolutional encoding layer and decoding layer. Secondly, the training data samples are imported, the convolutional model is compiled, and the network structure parameters are tuned. The parameters of the one-dimensional convolutional autoencoder network are finally determined as shown in Table 1. Its structure includes an input layer, two convolutional pooling layers, two deconvolutional pooling layers and an output layer. During the experiment, the two-layer network can obtain the abstract features of the sample data. Since the data itself has low noise, the convolution kernel size in the network is selected to be small, and the network has good feature extraction capabilities.

[0057] Table 1 Network structure parameters

[0058]

[0059]

[0060] The encoding feature extraction layer of 1DCAE is an important stage of autoencoding. In order to gain a deeper understanding of the feature extraction effect of 1DCAE and its effectiveness in extracting transient fault data features, a 1DCAE feature visualization network is constructed. Figure 4 As shown in the figure, the six types of fault data in the sample data are visualized respectively. First, the principal component analysis (PCA) is performed on the 768-dimensional original data of each type of data, and the original data is reduced to 2 dimensions by PCA technology. It can be seen that the original data is scattered and the clustering characteristics of the same type of label data are not obvious; secondly, the original data is imported into the one-dimensional convolutional autoencoder network to extract low-dimensional features, and then the PCA technology is used to reduce it to 2 dimensions. Compared with the 2-dimensional visualization features of the original data, it can be seen that the encoded data features are more regular than the original data distribution, the same type of label data is clustered, and the characteristics of various types of faults are more prominent. Selecting data with more prominent fault characteristics for back-end classification and recognition can not only improve the network's computing efficiency but also improve the network's recognition accuracy.

[0061] 2) Research on fault data classification

[0062] Based on the same experimental environment, a 1DCNN classification framework was built and data was imported for training. The parameters of 1DCNN are shown in Table 2. It includes an input layer, three convolutional layers, three pooling layers, and an output layer. The original 768-dimensional data was divided into a training set and a test set in a ratio of 8:2 for network training. The loss function of the training process is shown in Figure 5(a), it can be seen that when the number of training times reaches 350, the mean square error has stabilized and the network structure parameters have stabilized, so 350 is selected as the training stop condition. The accuracy of the training process is shown in Figure 5 (b) When the number of training times reaches 96, the accuracy rate exceeds 87%. As the number of training iterations increases, the accuracy rate increases slowly, and the final classification accuracy rate of the original data reaches 88.6%. The original data is extracted through the 1DCAE network to extract the low-dimensional features of the fault data, and the reduced-dimensional feature data is divided into the training set and the test set in the same proportion and fed into the 1DCNN, that is, the back-end convolutional neural network is used to classify the low-dimensional feature data. The recognition accuracy of the training process is shown in Figure 5 (b) It can be seen that when the number of training times reaches 180, the recognition accuracy has exceeded 91%. As the number of training times increases, the final recognition accuracy of the convolutional neural network reaches 93.75%. Comparing the training process parameters of the two models, it can be seen that in the early stage of training (the number of training times is between 0 and 161 times), the 1DCNN model converges faster and the recognition accuracy is higher than that of the 1DCAE-1DCNN model; in the late stage of training (the number of training times is between 161 and 350 times), the 1DCAE-1DCNN model converges significantly faster than the 1DCNN model, and the recognition accuracy exceeds that of the 1DCNN model.

[0063] Table 2 Network structure parameters

[0064]

[0065] 3) Comparative experimental analysis

[0066] Based on the same experimental environment and the same data samples, the end-to-end dual convolutional neural network classification model is compared with traditional classification methods such as decision tree, SVM, ANN, and SAE-CNN models.

[0067] Table 3 Comparison of recognition accuracy of different models

[0068]

[0069] As shown in Table 3, the statistical results of the recognition accuracy of the model of the present invention and decision tree, SVM, ANN, CNN, and SAE-CNN for the test data are shown. It can be seen that the recognition rate of transient fault data by the traditional machine learning model is low; the use of convolutional neural network in the model can improve the recognition rate of complex data samples, and the recognition rate is above 88%; adding a feature extraction network to the front end of the classification model can improve the recognition accuracy of the model; among different classification models, the recognition accuracy of 1DCAE-1DCNN for transient fault data is significantly better than other models, and its recognition accuracy reaches 93.8%.

[0070] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A distribution network fault data identification method based on convolutional neural network, characterized in that: The method specifically comprises the following steps: S1: Obtain transient fault data of distribution network and pre-process it; S2: Use multi-layer 1DCAE to build a fault feature extraction network, and input the pre-processed transient fault data to train and optimize the fault feature extraction network to obtain low-dimensional fault features, and then reduce the extracted low-dimensional fault features to two dimensions through principal component analysis; 1DCAE represents a one-dimensional convolutional autoencoder; Specifically, a fault feature extraction network is built by stacking multiple layers of 1DCAE, and different types of preprocessed transient fault data are uniformly used for network encoding and decoding training. The network parameters are adjusted to optimize the network weights and biases so that the network has the best feature extraction effect. Then, low-dimensional fault features are output from the encoder for fault data classification and identification in step S3. The 1DCAE is composed of a convolution layer, a pooling layer, a de-pooling layer and a deconvolution layer; the convolutional pooling data is reconstructed by deconvolution and de-pooling using the spatial information preservation characteristics of convolution, and then the model is solved by optimizing the minimum error between the reconstructed data and the original data, and finally the encoded low-dimensional effective features are extracted; S3: Use multi-layer 1DCNN to build a fault feature classification model, and input low-dimensional fault features to train and optimize the fault feature classification model to complete the recognition of different types of transient fault data; 1DCNN represents a one-dimensional convolutional neural network; Specifically, the low-dimensional fault features obtained in step S2 are relearned to train the optimal parameters of the network; then the fault features are classified through the convolutional network Softmax classifier layer to complete the recognition of different types of transient fault data; The 1DCNN consists of a one-dimensional convolutional layer, a one-dimensional pooling layer, a fully connected layer and a classification layer, and cooperates with activation functions, regularization, and optimizers to complete training model tuning; the fully connected layer is a data block after convolution and pooling, which is converted into a one-dimensional fully connected layer. The dropout method is added in front of it to randomly delete some neurons to suppress model overfitting, and the multi-classification of data can be completed in conjunction with the classifier.

2. The method for identifying distribution network fault data according to claim 1, characterized in that: Step S1 specifically includes: obtaining the electrical quantity information of the fault line recorded in the massive wave recording file of the master station, that is, the transient fault data of the distribution network; the preprocessing stage is to intercept the electrical quantity information near the fault point that best reflects the fault characteristics as the training data sample of the network.

3. The method for identifying distribution network fault data according to claim 2, characterized in that: In step S1, the preprocessing specifically includes: the acquired distribution network transient fault data is composed of 256 three-phase voltage sampling data before and after the fault moment; and the sample is expanded.

4. The method for identifying distribution network fault data according to claim 1, characterized in that: In step S3, the output of the fully connected layer is: in, is the output of the fully connected layer, is the fully connected layer input, ω is the weight, is the bias, f(·) is the activation function, and l is the number of fully connected layers; The classifier uses the Softmax classifier with a fully connected layer to complete data multi-classification, and the expression is: Among them, y j is the jth output neuron, and n is the number of neurons.

Images