Low-voltage distribution network single-phase earth fault line selection method based on deep transfer learning

Abstract

The invention belongs to the technical field of power distribution, and particularly relates to a low-voltage power distribution network single-phase earth fault line selection method based on deep transfer learning. The method comprises the following steps: calculating a characteristic value of transient fault component data of each phase of current in a time domain; based on the characteristic value of the transient fault component data of each phase of current, an automatic encoder-classifier fusion neural network suitable for low-voltage power distribution network single-phase earth fault line selection is constructed, and the fusion neural network comprises an encoder part of an automatic encoder and a classifier; training the automatic encoder by using a data set formed by single-phase earth fault data under different fault conditions; migrating the encoder part of the trained automatic encoder from a source domain to a target domain by using deep migration learning, and constructing a fusion neural network with the classifier; and adaptively adjusting the number of layers and the weight of the neural network by using an iteration loss function. According to the invention, effective fault line selection can be realized under the conditions of insufficient fault data and unobvious fault features.

Description

Technical Field

[0001] The present invention belongs to the technical field of power distribution, and in particular relates to a single-phase grounding fault line selection method for a low-voltage distribution network based on deep transfer learning. Background Art

[0002] my country has the world's largest power distribution system, which is mainly composed of 3-35KV voltage level distribution networks. Most of these distribution networks are small current grounding systems. According to the specific form of neutral point grounding, the system can be further divided into two types: ungrounded and grounded through arc suppression coils. According to relevant data from the State Grid, in my country, more than 80% of the faults in the small current grounding system are single-phase grounding faults. Single-phase grounding faults in distribution networks have brought great troubles to the safe use of electricity in life. Therefore, it is of great practical significance to study accurate and reliable line selection technology and troubleshoot faults in a timely manner.

[0003] The processing technology of low-voltage distribution network after single-phase grounding fault has been one of the long-standing problems, especially the arc suppression coil grounding system. For the single-phase grounding fault line selection problem, most traditional methods extract the characteristics of the signal and compare it with a set threshold to determine the fault line. Traditional fault line selection methods include external signal method, fault steady-state signal method, fault transient signal method, and data-driven line selection method.

[0004] In traditional methods, the signal injection method mostly requires the transformation of the primary system, which is difficult to implement. The fault steady-state signal method is limited by the weak fault characteristics in high-resistance faults, and the reliability and sensitivity of line selection cannot be guaranteed. The transient signal method uses the extracted signal characteristics to compare with the setting threshold to determine the faulty feeder. In the actual environment, due to the complex on-site working conditions, grounding medium, and frequent operation mode adjustments, it may lack adaptability and require dynamic adjustment of the threshold. The data-driven method needs to obtain a variety of fault training samples, extract the characteristic information of the fault through a large number of training samples, and does not rely on a single threshold criterion. In particular, methods such as deep neural networks have high adaptability and meet the complex and changeable distribution network. It is a more effective method. However, it is difficult to obtain a large number of labeled training samples in the actual environment, making it difficult for the existing methods to achieve the best effect.

[0005] Therefore, a neural network model is constructed based on the automatic encoder and encoder based on 9 eigenvalues ​​such as mean value, standard deviation, skewness factor, maximum value, minimum value, peak-to-peak value, root mean square, amplitude factor and waveform factor to realize intelligent line selection for single-phase grounding fault in low-voltage distribution network, which has good practical engineering application value. Summary of the invention

[0006] The purpose of the present invention is to address the above technical problems and propose a single-phase grounding fault line selection method for low-voltage distribution network based on deep transfer learning.

[0007] In order to achieve the above-mentioned purpose, the present invention adopts the following technical solutions:

[0008] The single-phase ground fault line selection method for low-voltage distribution network based on deep transfer learning includes the following steps:

[0009] S1: Collect transient three-phase currents before and after a fault occurs in the distributed low-voltage distribution network on the busbar and each feeder of the low-voltage distribution network, and obtain the transient fault component data of the phase current of the three-phase current according to the collected transient three-phase currents before and after a single-phase grounding fault occurs in the distribution network;

[0010] S2: Calculate the characteristic values ​​of the transient fault component data of each phase current obtained in S1 in the time domain, wherein the characteristic values ​​include the mean value, standard deviation, skewness factor, maximum value, minimum value, peak-to-peak value, root mean square, crest factor, and waveform factor, a total of 9-dimensional characteristic values;

[0011] S3: Based on the transient three-phase current obtained by S1 and the characteristic values ​​of the transient fault component data of each phase current calculated by S2, an autoencoder-classifier fusion neural network suitable for single-phase grounding fault in the distribution network is constructed. The neural network fuses the encoder part of the autoencoder and the classifier network model by deep transfer learning, and then trains the neural network using a data set consisting of single-phase grounding fault data under different fault conditions of the fault feeder of the distribution network;

[0012] S4: When training the constructed autoencoder-classifier fusion neural network, the fault data set is used as the input, written into the neural network model, and iterative calculation is performed to minimize the gap between the output result of the trained fusion neural network model and the input data; the number of layers and weights of the neural network model are optimized so that the fusion neural network can achieve accurate and reliable single-phase grounding fault line selection.

[0013] Further, as a preferred technical solution of the present invention, S3 specifically includes the following steps:

[0014] S3.1. Construct an autoencoder based on 9 eigenvalues, namely, mean value, standard deviation, skewness factor, maximum value, minimum value, peak-to-peak value, root mean square, crest factor, and waveform factor. The input layer has k units and the output layer also has k units, where k is the total number of distribution network buses and outgoing feeders.

[0015] S3.2. In a given distribution network simulation model, single-phase grounding faults under different fault conditions are set on k distribution network buses and outgoing feeders in turn, and a total of M fault data samples are obtained as a data set;

[0016] S3.3, divide the fault data samples into p1 training samples, q1 test samples and r1 verification samples, r1<q1<p1<M, p1+q1+r1=M; use p1 training samples to train the autoencoder neural network constructed in S3.1, after the neural network is trained, test it with q1 test samples, optimize the number of layers of the neural network and adjust its weights according to the test results, and finally verify the constructed neural network with r1 verification samples;

[0017] S3.4, the encoder part of the autoencoder trained in S3.3 is transferred from the source domain to the target domain by the deep transfer learning method, and a fusion neural network for single-phase grounding fault line selection in the distribution network is constructed with the classifier; the autoencoder-classifier fusion neural network includes an input layer, a convolution layer, a pooling layer, a fully connected layer, a softmax layer and an output layer; the signal is mapped after different convolution layers and pooling layers, and the extracted features are classified in the fully connected layer and the softmax layer, and finally a new expression is output; specifically:

[0018]

[0019] In the formula, is the output and input of the jth layer, y k is the kth output, y g is the g-th value of the output layer, is the convolution kernel, w k 、w g is the weight, b z 、b k 、b g is the bias; f is the neuron activation function of the neural network; dwon is the downsampling function of the pooling layer;

[0020] S3.5. Divide the fault data samples into p2 training samples, q2 test samples and r2 verification samples, r2<q2<p2<M, p2+q2+r2=M. Use p2 training samples to train the distribution network single-phase grounding fault line selection fusion neural network constructed in S3.4. After the neural network is trained, it is tested with q2 test samples. According to the probability of the faulty line and the probability of the sound line in the test results, the gradient descent algorithm is used to optimize the number of layers of the neural network and fine-tune the weights therein. Finally, the constructed neural network is verified through r2 verification samples.

[0021] Further as a preferred technical solution of the present invention,

[0022] The deep transfer learning described in S3.4 specifically includes: using the target data and the source data to obtain the regression coefficient vector of the transfer learning L(ω;A(e) ,B (e) ) is the loss function of the model; the target model of the fusion neural network is

[0023]

[0024] Finally, based on the source domain task Ts and the target domain task Tt, the source domain Ds and the target domain Dt are continuously optimized to achieve migration;

[0025] Ds, Ts and are the source domain, source domain task and source data of the autoencoder, Dt, Tt and (A (0) ,B (0) ) are the target domain, target domain task and target data of the encoder-classifier fusion neural network respectively; argmin is to take the parameter value that minimizes the loss function of the model in its definition domain; LOGIT is the inverse transform of the sigmoid function, in order to make the final probability between 0 and 1.

[0026] Further, as a preferred technical solution of the present invention, the different fault conditions in S3.2 include: different transition resistances on the feeder outgoing from the busbar, different fault initial phase angles, and different fault occurrence locations on the line.

[0027] Further, as a preferred technical solution of the present invention, the S4 specifically includes the following steps:

[0028] S4.1, the zero-sequence current of the busbar and its outgoing feeder together constitute X, X∈Rd represents the input of the neural network model; the data label c={0,1,2,…,n}, the training sample set D={(xi,ci)|1≤i≤s}, where xi∈X, ci∈c, s is the total number of samples in the training sample set; the neural network model is equivalent to a multi-label classifier model, and its output is an n-dimensional vector, represented as {m1,…,mn}; based on softmax regression, the probability that the input X belongs to the i-th label information is

[0029] S4.2, set the filter, convolution kernel size, convolution step size, use ReLu as the activation function, set the pooling window size, set the number of data samples captured for one training and the number of iterations;

[0030] S4.3. Based on the probability Q of S4.1, we get the loss function of the fusion neural network Through iteration and adjustment of weights, the optimal parameters are obtained, so that the gap between the output result of the fusion neural network and the input data is minimized and the line selection accuracy is maximized.

[0031] The method for selecting a single-phase ground fault line in a low-voltage distribution network based on deep transfer learning according to the present invention adopts the above technical solution and has the following technical effects compared with the prior art:

[0032] (1) According to the amplitude and volatility indicators of the transient current of the single-phase grounding fault in the distribution network in the time domain, the present invention first selects the characteristic values ​​of the transient fault component data of each phase current (such as the mean value, standard deviation, skewness factor, maximum value, minimum value, peak-to-peak value, root mean square, amplitude factor, and waveform factor, a total of 9 dimensional characteristic values), and then constructs a fusion neural network for single-phase grounding fault line selection in the distribution network composed of the encoder part of the automatic encoder and the classifier.

[0033] (2) Through the unsupervised learning of the autoencoder and the transfer learning method, the constructed fusion neural network can still maintain accurate and reliable fault line selection when faced with small data amounts and insufficient data labels.

[0034] (3) Deep learning can achieve deeper mining of data information and mapping relationships. Combined with the fault data accumulated by the distribution system data acquisition device, through data set construction and network training and learning, the type and location of distribution network faults can be identified, thereby improving line selection efficiency and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is an autoencoder neural network model with deep transfer learning capability according to an embodiment of the present invention;

[0036] Figure 2 The encoder-classifier fusion neural network model for line selection of single-phase grounding fault in distribution network according to an embodiment of the present invention;

[0037] Figure 3 It is a structural diagram of a single-phase grounding fault simulation model of a distribution network according to an embodiment of the present invention. DETAILED DESCRIPTION

[0038] The present invention is further explained below in detail with reference to the accompanying drawings so that those skilled in the art can more deeply understand the present invention and be able to implement it. However, the following reference examples are only used to explain the present invention and are not intended to limit the present invention.

[0039] The single-phase ground fault line selection method for low-voltage distribution network based on deep transfer learning includes the following steps:

[0040] S1: Collect transient three-phase currents before and after a fault occurs in the distributed low-voltage distribution network on the busbar and each feeder of the low-voltage distribution network, and obtain the transient fault component data of the phase current of the three-phase current according to the collected transient three-phase currents before and after a single-phase grounding fault occurs in the distribution network;

[0041] S2: Calculate the characteristic values ​​of the transient fault component data of each phase current obtained in S1 in the time domain, wherein the characteristic values ​​include the mean value, standard deviation, skewness factor, maximum value, minimum value, peak-to-peak value, root mean square, crest factor, and waveform factor, a total of 9-dimensional characteristic values;

[0042] S3: Based on the transient three-phase current obtained by S1 and the characteristic values ​​of the transient fault component data of each phase current calculated by S2, an autoencoder-classifier fusion neural network suitable for single-phase grounding fault in the distribution network is constructed. The neural network fuses the encoder part of the autoencoder and the classifier network model by deep transfer learning, and then trains the neural network using a data set consisting of single-phase grounding fault data under different fault conditions of the fault feeder of the distribution network;

[0043] S4: When training the constructed autoencoder-classifier fusion neural network, the fault data set is used as the input, written into the neural network model, and iterative calculation is performed to minimize the gap between the output result of the trained fusion neural network model and the input data; the number of layers and weights of the neural network model are optimized so that the fusion neural network can achieve accurate and reliable single-phase grounding fault line selection.

[0044] S3 specifically includes the following steps:

[0045] S3.1. Construct an autoencoder based on 9 eigenvalues, namely, mean value, standard deviation, skewness factor, maximum value, minimum value, peak-to-peak value, root mean square, crest factor, and waveform factor. The input layer has k units and the output layer also has k units, where k is the total number of distribution network buses and outgoing feeders.

[0046] S3.2. In a given distribution network simulation model, single-phase grounding faults under different fault conditions are set on k distribution network buses and outgoing feeders in turn, and a total of M fault data samples are obtained as a data set;

[0047] S3.3, divide the fault data samples into p1 training samples, q1 test samples and r1 verification samples, r1<q1<p1<M, p1+q1+r1=M; use p1 training samples to train the autoencoder neural network constructed in S3.1, after the neural network is trained, test it with q1 test samples, optimize the number of layers of the neural network and adjust its weights according to the test results, and finally verify the constructed neural network with r1 verification samples;

[0048] S3.4, the encoder part of the autoencoder trained in S3.3 is transferred from the source domain to the target domain by the deep transfer learning method, and a fusion neural network for single-phase grounding fault line selection in the distribution network is constructed with the classifier; the autoencoder-classifier fusion neural network includes an input layer, a convolution layer, a pooling layer, a fully connected layer, a softmax layer and an output layer; the signal is mapped after different convolution layers and pooling layers, and the extracted features are classified in the fully connected layer and the softmax layer, and finally a new expression is output; specifically:

[0049]

[0050] In the formula, is the output and input of the jth layer, y k is the kth output, y g is the g-th value of the output layer, is the convolution kernel, w k 、w g is the weight, b z 、b k 、b g is the bias; f is the neuron activation function of the neural network; dwon is the downsampling function of the pooling layer;

[0051] S3.5. Divide the fault data samples into p2 training samples, q2 test samples and r2 verification samples, r2<q2<p2<M, p2+q2+r2=M. Use p2 training samples to train the distribution network single-phase grounding fault line selection fusion neural network constructed in S3.4. After the neural network is trained, it is tested with q2 test samples. According to the probability of the faulty line and the probability of the sound line in the test results, the gradient descent algorithm is used to optimize the number of layers of the neural network and fine-tune the weights therein. Finally, the constructed neural network is verified through r2 verification samples.

[0052] The deep transfer learning described in S3.4 specifically includes: using the target data and the source data to obtain the regression coefficient vector of the transfer learning L(ω;A (e) ,B (e) ) is the loss function of the model; the target model of the fusion neural network is

[0053]

[0054] Finally, based on the source domain task Ts and the target domain task Tt, the source domain Ds and the target domain Dt are continuously optimized to achieve migration;

[0055] Ds, Ts and are the source domain, source domain task and source data of the autoencoder, Dt, Tt and (A (0) ,B(0) ) are the target domain, target domain task and target data of the encoder-classifier fusion neural network respectively; argmin is to take the parameter value that minimizes the loss function of the model in its definition domain; LOGIT is the inverse transform of the sigmoid function, in order to make the final probability between 0 and 1.

[0056] The different fault conditions in S3.2 include: different transition resistances on the feeders from the busbar, different fault initial phase angles, and different fault locations on the line.

[0057] S4 specifically includes the following steps:

[0058] S4.1, the zero-sequence current of the busbar and its outgoing feeder together constitute X, X∈Rd represents the input of the neural network model; the data label c={0,1,2,…,n}, the training sample set D={(xi,ci)|1≤i≤s}, where xi∈X, ci∈c, s is the total number of samples in the training sample set; the neural network model is equivalent to a multi-label classifier model, and its output is an n-dimensional vector, represented as {m1,…,mn}; based on softmax regression, the probability that the input X belongs to the i-th label information is

[0059] S4.2, set the filter, convolution kernel size, convolution step size, use ReLu as the activation function, set the pooling window size, set the number of data samples captured for one training and the number of iterations;

[0060] S4.3. Based on the probability Q of S4.1, we get the loss function of the fusion neural network Through iteration and adjustment of weights, the optimal parameters are obtained, so that the gap between the output result of the fusion neural network and the input data is minimized and the line selection accuracy is maximized.

[0061] When implementing it, Figure 1 As shown, the present invention includes: step 1: collecting transient three-phase currents before and after a fault occurs in a distributed low-voltage distribution network, and obtaining transient fault component data of phase currents of the three-phase currents according to the collected transient three-phase currents before and after a single-phase grounding fault occurs in the distribution network. At the transient recording terminal, when a single-phase grounding fault occurs in the distribution network, a zero-sequence current mutation is monitored at the bus to determine the time when the fault occurs.

[0062] Step 2: In the time domain, calculate the characteristic values ​​of the transient fault component data of each phase current obtained in step 1, the characteristic values ​​include mean value, standard deviation, skewness factor, maximum value, minimum value, peak-to-peak value, root mean square, crest factor, and waveform factor, a total of 9-dimensional characteristic values, and the specific calculation formulas are as follows:

[0063] average value

[0064] Standard Deviation

[0065] Skewness factor

[0066] Maximum value X4 = maxx i

[0067] Minimum value X5 = minx i

[0068] Peak-to-peak value X6 = maxx i -minx i

[0069] RMS

[0070] Crest factor X8 = X6 / X7

[0071] Form Factor

[0072] Step 3: Based on the transient three-phase current obtained in step 1 and the characteristic values ​​of the transient fault component data of each phase current calculated in step 2, an automatic encoder-classifier fusion neural network suitable for single-phase grounding fault in the distribution network is constructed, and then the automatic encoder-classifier fusion neural network is trained using a data set consisting of single-phase grounding fault data with different transition resistances, different fault initial phase angles, and different fault locations of the fault feeder in the distribution network.

[0073] Since the fault characteristics of single-phase grounding fault in distribution network are complex and changeable, in order to alleviate the problem of unclear fault characteristics and less actual fault data, an encoder-classifier fusion neural network suitable for line selection of single-phase grounding fault in distribution network is constructed. Figure 2 shown.

[0074] Step 3 specifically includes:

[0075] Step 3.1: Construct an autoencoder based on 9 eigenvalues, namely, mean value, standard deviation, skewness factor, maximum value, minimum value, peak-to-peak value, root mean square, crest factor, and waveform factor. The input layer has k units and the output layer also has k units, where k is the total number of distribution network buses and outgoing feeders.

[0076] Step 3.2: If Figure 3 As shown, in a given distribution network simulation model, single-phase grounding faults with different transition resistances, different fault initial phase angles, and different fault locations are set on k distribution network buses and their outgoing feeders in turn to obtain a total of M fault data samples.

[0077] Step 3.3: Divide the fault data samples into 70% training samples, 20% test samples and 10% validation samples. Use the 70% training samples to train the autoencoder neural network constructed in step 3.1. After the neural network is trained, it is tested with 20% test samples. According to the test results, the number of layers of the neural network is optimized and its weights are adjusted. Finally, the constructed neural network is verified with 10% validation samples.

[0078] Step 3.4: The encoder part of the autoencoder trained in step 3.3 is transferred from the source domain to the target domain through the deep transfer learning method, and the fusion neural network for single-phase grounding fault line selection in the distribution network is constructed with the classifier. The source domain Ds and source domain task Ts of the encoder part of the autoencoder, the target domain Dt and target domain task Tt of the encoder-classifier fusion neural network. Based on the source domain task Ts and the target domain task Tt, the source domain Ds and the target domain Dt are continuously optimized to achieve migration.

[0079] Step 3.5: Divide the fault data samples into 70% training samples, 20% test samples and 10% verification samples. Use the 70% training samples to train the distribution network single-phase grounding fault line selection fusion neural network constructed in step 3.4. After the neural network is trained, it is tested with 20% test samples. According to the test results, the real probability of the faulty feeder in the sample is not 100% and the probability of other feeders is not 0%. Feedback is used to optimize the number of layers of the neural network and fine-tune the weights therein using the gradient descent algorithm. Finally, the constructed neural network is verified by 10% verification samples.

[0080] When a single-phase grounding fault occurs in a low-voltage distribution network, the fault characteristics are not obvious. Deep learning can achieve deeper mining of data information and the mapping relationship between them, thereby improving the efficiency and accuracy of line selection.

[0081] Due to the different fault probabilities of each feeder in the actual environment, there is a more obvious imbalance in the label samples. The autoencoder has the ability to learn without supervision without labels, and combines the trained encoder with the classifier neural network through transfer learning. This improves the accuracy of line selection for single-phase grounding faults in low-voltage distribution networks.

[0082] The experimental results prove that the scheme can automatically process the data characteristics of single-phase grounding faults in low-voltage distribution networks, eliminating the process of manual labeling and realizing autonomous, efficient and high-precision fault line selection in low-voltage distribution networks.

[0083] The specific implementation scheme described above further describes in detail the purpose, technical scheme and beneficial effects of the present invention. It should be understood that the above is only a specific implementation scheme of the present invention and is not intended to limit the scope of the present invention. Any equivalent changes and modifications made by any technician in the field without departing from the concept and principle of the present invention should fall within the scope of protection of the present invention.

Claims

1. A single-phase ground fault line selection method for low-voltage distribution network based on deep transfer learning, characterized in that: The following steps are involved: S1: Collect transient three-phase currents before and after a fault occurs in the distributed low-voltage distribution network on the busbar and each feeder of the low-voltage distribution network, and obtain the transient fault component data of the phase current of the three-phase current according to the collected transient three-phase currents before and after a single-phase grounding fault occurs in the distribution network; S2: Calculate the characteristic values ​​of the transient fault component data of each phase current obtained in S1 in the time domain, wherein the characteristic values ​​include the mean value, standard deviation, skewness factor, maximum value, minimum value, peak-to-peak value, root mean square, crest factor, and waveform factor, a total of 9-dimensional characteristic values; S3: Based on the transient three-phase current obtained by S1 and the characteristic values ​​of the transient fault component data of each phase current calculated by S2, an autoencoder-classifier fusion neural network suitable for single-phase grounding fault in the distribution network is constructed. The neural network fuses the encoder part of the autoencoder and the classifier network model by deep transfer learning, and then trains the neural network using a data set consisting of single-phase grounding fault data under different fault conditions of the fault feeder of the distribution network; S4: When training the constructed autoencoder-classifier fusion neural network, the fault data set is used as the input, written into the neural network model, and iterative calculation is performed to minimize the gap between the output result of the trained fusion neural network model and the input data; the number of layers and weights of the neural network model are optimized so that the fusion neural network can achieve accurate and reliable single-phase grounding fault line selection.

2. The method for selecting a single-phase ground fault line in a low-voltage distribution network based on deep transfer learning according to claim 1 is characterized in that: The S3 specifically includes the following steps: S3.

1. Construct an autoencoder based on 9 eigenvalues, namely, mean value, standard deviation, skewness factor, maximum value, minimum value, peak-to-peak value, root mean square, crest factor, and waveform factor. The input layer has k units and the output layer also has k units, where k is the total number of distribution network buses and outgoing feeders. S3.

2. In a given distribution network simulation model, single-phase grounding faults under different fault conditions are set on k distribution network buses and outgoing feeders in turn, and a total of M fault data samples are obtained as a data set; S3.3, divide the fault data samples into p1 training samples, q1 test samples and r1 verification samples, r1<q1<p1<M, p1+q1+r1=M; use p1 training samples to train the autoencoder neural network constructed in S3.1, after the neural network is trained, test it with q1 test samples, optimize the number of layers of the neural network and adjust its weights according to the test results, and finally verify the constructed neural network with r1 verification samples; S3.4, the encoder part of the autoencoder trained in S3.3 is transferred from the source domain to the target domain by the deep transfer learning method, and a fusion neural network for single-phase grounding fault line selection in the distribution network is constructed with the classifier; the autoencoder-classifier fusion neural network includes an input layer, a convolution layer, a pooling layer, a fully connected layer, a softmax layer and an output layer; the signal is mapped after different convolution layers and pooling layers, and the extracted features are classified in the fully connected layer and the softmax layer, and finally a new expression is output; specifically: In the formula, is the output and input of the jth layer, y k is the kth output, y g is the g-th value of the output layer, is the convolution kernel, w k 、w g is the weight, b z 、b k 、b g is the bias; f is the neuron activation function of the neural network; dwon is the downsampling function of the pooling layer; S3.

5. Divide the fault data samples into p2 training samples, q2 test samples and r2 verification samples, r2<q2<p2<M, p2+q2+r2=M. Use p2 training samples to train the distribution network single-phase grounding fault line selection fusion neural network constructed in S3.

4. After the neural network is trained, it is tested with q2 test samples. According to the probability of the faulty line and the probability of the sound line in the test results, the gradient descent algorithm is used to optimize the number of layers of the neural network and fine-tune the weights therein. Finally, the constructed neural network is verified through r2 verification samples.

3. The method for selecting a single-phase ground fault line in a low-voltage distribution network based on deep transfer learning according to claim 2 is characterized in that: The deep transfer learning described in S3.4 specifically includes: using the target data and the source data to obtain the regression coefficient vector of the transfer learning e=0,···,E; L(ω;A (e) ,B (e) ) is the loss function of the model; the target model of the fusion neural network is Finally, based on the source domain task Ts and the target domain task Tt, the source domain Ds and the target domain Dt are continuously optimized to achieve migration; Ds, Ts and are the source domain, source domain task and source data of the autoencoder, Dt, Tt and (A (0) ,B (0) ) are the target domain, target domain task and target data of the encoder-classifier fusion neural network respectively; argmin is to take the parameter value that minimizes the loss function of the model in its definition domain; LOGIT is the inverse transform of the sigmoid function, in order to make the final probability between 0 and 1.

4. The method for selecting a single-phase ground fault line in a low-voltage distribution network based on deep transfer learning according to claim 2 is characterized in that: The different fault conditions in S3.2 include: different transition resistances on the feeders from the busbar, different fault initial phase angles, and different fault locations on the line.

5. The method for selecting a single-phase ground fault line in a low-voltage distribution network based on deep transfer learning according to claim 2 is characterized in that: The S4 specifically comprises the following steps: S4.1, the zero-sequence current of the busbar and its outgoing feeder together constitute X, X∈Rd represents the input of the neural network model; the data label c={0,1,2,…,n}, the training sample set D={(xi,ci)|1≤i≤s}, where xi∈X, ci∈c, s is the total number of samples in the training sample set; the neural network model is equivalent to a multi-label classifier model, and its output is an n-dimensional vector, represented as {m1,…,mn}; based on softmax regression, the probability that the input X belongs to the i-th label information is S4.2, set the filter, convolution kernel size, convolution step size, use ReLu as the activation function, set the pooling window size, set the number of data samples captured for one training and the number of iterations; S4.

3. Based on the probability Q of S4.1, we get the loss function of the fusion neural network Through iteration and adjustment of weights, the optimal parameters are obtained, so that the gap between the output result of the fusion neural network and the input data is minimized and the line selection accuracy is maximized.

Images