A method for detecting partial discharge in power transmission lines

Abstract

This invention relates to a method for detecting partial discharge in power transmission lines, comprising the following steps: S1, acquiring raw electrical signals from the power transmission line and dividing them into training samples and test samples; S2, batch inputting the training samples into a multi-scale one-dimensional convolutional neural network, setting thresholds and the structure and kernel size of the multi-scale one-dimensional convolutional neural network, propagating forward layer by layer to obtain the features of the training samples, and calculating the error between the expected output and the actual output; S3, propagating the error backward from the output to the input, fine-tuning the network parameters layer by layer; S4, repeating steps S2 and S3 until all training samples are completed, calculating the network's accuracy on the test samples, if the accuracy is greater than the set threshold, an ideal network is obtained; otherwise, returning to step S2; S5, outputting the accuracy of the ideal network on the test samples. This invention solves the problem of partial discharge diagnosis in power transmission lines, achieving good results under different frequency data and different convolution kernel settings.

Description

Technical Field

[0001] This invention relates to the field of power distribution network fault diagnosis, specifically to a method for detecting partial discharge in transmission lines based on a multi-scale one-dimensional convolutional neural network. Background Technology

[0002] In recent years, the widespread integration of distributed power sources into new power systems has led to increasingly complex power grid topologies. Consequently, the causes of transmission line faults are difficult to analyze. Partial discharge is a common fault in power systems, but it does not cause immediate power outages for users; therefore, manual inspection of long-distance transmission lines is costly. Transmission lines are a crucial link in ensuring power supply reliability, and their insulation diagnosis is of great significance, with partial discharge detection being a necessary step in insulation diagnosis. Traditional partial discharge detection methods, such as acoustic, optical, and electromagnetic wave detection, mainly rely on feature extraction of fault signals, where the quality of feature extraction significantly impacts the detection results. With the rise of deep learning technology and the availability of large amounts of labeled data, automatic feature extraction of partial discharge has been applied in power systems. Automatic feature extraction and online identification of partial discharge are of great importance to automatic transmission line partial discharge detection systems.

[0003] Convolutional Neural Networks (CNNs) are an important component of computer vision and can also be applied to time series data. The difference lies in the kernel: convolutional kernels used for images are two-dimensional, while those used for time series data are one-dimensional, hence the term "one-dimensional convolutional neural network." Compared to models like Long Short-Term Memory (LSTM) recurrent neural networks (RNNs), one-dimensional CNNs offer advantages such as faster training speeds and parallel computation, and in certain scenarios, they can achieve performance comparable to LSTMs. Summary of the Invention

[0004] The purpose of this invention is to address the problem of partial discharge diagnosis in power transmission lines by proposing a method for detecting partial discharge in power transmission lines. This method mainly employs a multi-scale one-dimensional convolutional neural network model to achieve partial discharge detection in power transmission lines.

[0005] The model uses electrical signals of arbitrary length, which can be viewed as a binary classifier, making it easy to adjust the input sequence length according to the sampling rate of the diagnostic equipment. Partial discharge is an extreme event in the normal operation of a power system; therefore, the number of partial discharge fault samples will be much smaller than under normal conditions, posing a challenge to training a network to prevent overfitting. Therefore, it is necessary to select an appropriate proportion of partial discharge signals and normal signals as training data. The electrical signals of different lines vary greatly; therefore, we use a standardization method to adjust the input data to improve training accuracy. While these artificial data processing methods are not feature extraction methods, they are essential for model training.

[0006] This invention proposes a method for detecting partial discharge in power transmission lines, comprising the following steps:

[0007] S1. Obtain the original electrical signals of the transmission line and divide them into training samples and test samples;

[0008] S2. Input training samples into the multi-scale one-dimensional convolutional neural network in batches, set the desired accuracy threshold, set the structure of the scale one-dimensional convolutional neural network and its convolution kernel size, propagate forward layer by layer to obtain the features of the training samples, obtain the actual output, and calculate the error between the desired output and the actual output.

[0009] S3. The error is propagated back to the input layer by layer from the output of the multi-scale one-dimensional convolutional neural network in S1, updating the network parameters.

[0010] S4. Repeat steps S2 and S3 until all training samples are completed. Calculate the accuracy of the scale one-dimensional convolutional neural network on the test samples. If the accuracy is greater than the set threshold, the ideal network is obtained; otherwise, return to step S2.

[0011] S5. Output the accuracy of the ideal network on the test sample.

[0012] Furthermore, the original power transmission line electrical signal includes voltage measurement signals of the three-phase overhead power line, and the voltage measurement signal of each phase is performed simultaneously.

[0013] Furthermore, the multi-scale one-dimensional convolutional neural network comprises two multi-scale one-dimensional convolutional neural network layers, two fully connected layers, and finally outputs the result through the Softmax function.

[0014] Furthermore, each one-dimensional convolutional neural network uses convolutional kernels of different scales to perform convolution operations on input signals of different scales, thereby extracting signal features of different fineness.

[0015] Furthermore, S2 includes the following:

[0016] The original transmission line electrical signal has a single sample length of T and a phase of p, which is represented by a column vector: φ p [1:T]=[φ p,1 ;φ p,2 ;...,φ p,T ];

[0017] The max pooling layer performs subsampling with a stride of d. The input to the max pooling layer is X, and the output is Mp(X), where the t-th element of Mp(X) is Mp. t (X)=max(X d(t-1)+i ,1≤i≤a), where a is the pooling size;

[0018] The input of a one-dimensional convolutional neural network layer with ReLU activation is Y, and the output is represented as Conv(Y), where the t-th element of Conv(Y) is... Where c is the size of the convolution kernel, ω i b is the i-th weight of the kernel. t denoted as the deviation of the t-th element of the convolutional layer output; the ReLU activation function is: ReLU(x) = max(0,x), which is a non-linear mapping from neuron input to output, and its function operates by taking the maximum value between 0 and x.

[0019] Furthermore, the flattening layer connects the three-phase signals into a one-dimensional vector using the ReLU activation function; each neuron in the fully connected layer is fully connected to all neurons in the previous layer, so as to integrate local information according to different categories in the convolutional neural network layer or pooling layer.

[0020] Furthermore, the last layer of the multi-scale one-dimensional convolutional neural network is a Softmax function; since partial discharge detection is a binary classification problem, the Softmax function acts as a binary classifier, and its output can be calculated as follows:

[0021]

[0022] Among them, PD and The presence or absence of discharge is indicated in the output results, W and D are the weight matrix parameters and bias parameters, and Z is the input to the Softmax function.

[0023] Furthermore, if the input signal is a three-phase waveform, the output of the multi-scale convolutional neural network layer needs to be connected to a flattening layer to adjust the vector form of the fully connected layer.

[0024] This invention solves the problem of partial discharge diagnosis of transmission lines by constructing a power grid fault detection model based on a multi-scale one-dimensional convolutional neural network, and achieves good results under different frequency data and different convolutional kernel settings;

[0025] The multi-scale one-dimensional convolutional neural network model proposed in this invention is suitable for the diagnosis of partial discharge signals in single-phase and three-phase circuits. This model does not require specifying the length of the input sequence, nor does it require manually extracting the features of the input signal. Attached Figure Description

[0026] Figure 1 This is a flowchart of the partial discharge detection method for power transmission lines according to the present invention;

[0027] Figure 2 The dataset in this embodiment of the invention contains the time-series waveform of phase A and related box plots.

[0028] Figure 3 The dataset in this embodiment of the invention contains the timing waveform of phase B and related box plots.

[0029] Figure 4 The dataset in this embodiment of the invention contains the time-series waveforms of phase C and related box plots.

[0030] Figure 5 This is a diagram of the multi-scale one-dimensional convolutional neural network structure according to an embodiment of the present invention;

[0031] Figure 6 This is a diagram of the multi-scale convolutional neural network layer structure according to an embodiment of the present invention;

[0032] Figure 7 The results show the comparison of partial discharge detection accuracy under different sampling frequencies (10MHz, 20MHz and 40MHz) and nine network architecture settings in this embodiment of the invention.

[0033] Figure 8 The kernel sizes for two types of multi-scale convolutional neural network layers;

[0034] Figure 9 Experimental setup for two combinations of multi-scale convolutional neural network layers;

[0035] Figure 10 ROC points for classifying positive samples with different probability thresholds. Detailed Implementation

[0036] The following detailed description, in conjunction with the accompanying drawings and specific embodiments, provides a further detailed explanation of the partial discharge detection method for power transmission lines proposed by the present invention. The advantages and features of the present invention will become clearer from the following description. It should be noted that the accompanying drawings are in a very simplified form and use non-precise proportions, used only to facilitate and clearly illustrate the embodiments of the present invention. Please refer to the accompanying drawings to make the objectives, features, and advantages of the present invention more apparent and understandable. It should be understood that the structures, proportions, sizes, etc., depicted in the accompanying drawings are only for illustrative purposes to aid those skilled in the art and are not intended to limit the implementation conditions of the present invention. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in proportions, or adjustments to the size, without affecting the effects and objectives achieved by the present invention, should still fall within the scope of the technical content disclosed in the present invention.

[0037] like Figure 1 As shown, the partial discharge detection method for power transmission lines proposed in this invention includes the following steps:

[0038] S1. Obtain the original electrical signals of the transmission line and divide them into training samples and test samples;

[0039] S2. Input training samples into the Multi-scale One Dimensional Convolutional Neural Network (MS-Conv1D) in batches, set a threshold, set the MS-Conv1D structure and its convolution kernel size, and extract the features of the training samples layer by layer in the forward propagation to obtain the actual output. Calculate the error between the expected output and the actual output.

[0040] S3. The error is propagated back to the input layer by layer from the output of MS-Conv1D in S1 to update the network parameters;

[0041] S4. Repeat steps S2 and S3 until all training samples are completed. Calculate the accuracy of the network on the test samples. If the accuracy is greater than the set threshold, the ideal network is obtained; otherwise, return to step S2.

[0042] S5. Output the accuracy of the ideal network on the test sample.

[0043] The raw transmission line electrical signals mentioned in step S1 are training set data obtained from the VSB's ENET center. This data was measured during grid operation in three-phase mode, with faulty phases marked during the measurements. The data includes voltage measurement signals for the three-phase overhead power lines, with each phase measured simultaneously. There are approximately 20k sets of single-phase voltage signals, or 7k complete three-phase measurement signals. Each signal set contains 800k sampling points, with a signal length equal to one periodic wave under a 50Hz grid, i.e., a sampling frequency of 40MHz. Figure 2 , Figure 3 and Figure 4 The waveforms and related box plots of phases A, B, and C with and without partial discharge are shown respectively.

[0044] As can be seen from the box plot, there are fewer voltage signal anomalies without partial discharge, and this characteristic can be used to classify partial discharge. However, when the local discharge effect is weak (e.g., Figure 4 For voltage signals containing partial discharge, this direct approach may fail. Therefore, automatic feature extraction deserves more attention.

[0045] From the perspective of learning local receptive field feature information, MS-Conv1D described in step S2 allows convolutional kernels of different scales to extract original power transmission line electrical signal features of varying fineness. Smaller kernels extract more detailed signal features and are more sensitive to high-frequency features. Larger kernels have a larger receptive field, enabling them to learn information over a longer time span, specifically the low-frequency features embedded in the original power transmission line electrical signal. Therefore, by combining the advantages of convolutional kernels of different scales and performing convolution operations with kernels of different scales, we can avoid ignoring signal features of varying fineness when using convolution at a single scale, thereby refining the signal features and allowing for a more comprehensive expression of the feature information.

[0046] The structure of MS-Conv1D described in step S2 is as follows: Figure 5As shown, single-phase or three-phase transmission line electrical signals are used as input, sequentially passing through two multi-scale one-dimensional convolutional neural network (CNN) layers, two fully connected layers, and finally outputting the judgment result through the Softmax function. If the input signal is a three-phase waveform, the output of the multi-scale CNN layer needs to be connected to a flattening layer to adjust the vector form of the fully connected layer. The multi-scale one-dimensional CNN layer is composed of a max-pooling layer 1, n parallel one-dimensional CNNs, a max-pooling layer 2, and a flattening layer connected sequentially and then concatenated. Each parallel one-dimensional CNN uses convolutional kernels of different scales to perform convolution operations on the original transmission line electrical signals of different scales, achieving the extraction of signal features of different fineness. Then, the features extracted by each one-dimensional CNN are concatenated and output to the max-pooling layer 2, as shown. Figure 6 As shown.

[0047] The multi-scale one-dimensional CNN submodule and training process described in step S2 are as follows: In S1, the length of a single sample of the original power transmission line electrical signal is T, and the phase is p, which can be represented by a column vector φ. p [1:T]=[φ p,1 ;φ p,2 ;...,φ p,T ].

[0048] The max pooling layer performs subsampling with a stride of d. The input to the max pooling layer is X, and the output is denoted by Mp(X), where the t-th element of Mp(X) is Mp. t (X)=max(X d(t-1)+i ,1≤i≤a), where a is the pooling size.

[0049] Similarly, a one-dimensional CNN with ReLU activation function takes Y as input and outputs Conv(Y), where the t-th element of Conv(Y) is... Where c is the size of the convolution kernel; ω i b is the i-th weight of the convolution kernel; t Let be the deviation of the t-th element of the convolutional layer output. The ReLU activation function is given by Relu(x) = max(0,x), which is a non-linear mapping from neuron input to output. The function operates by taking the maximum value between 0 and x.

[0050] The flattening layer connects the three-phase signals into a one-dimensional vector using the ReLU activation function. Each neuron in the fully connected layer is fully connected to all neurons in the previous layer to integrate local information according to different categories in the convolutional or pooling layers.

[0051] The last layer of the multi-scale one-dimensional convolutional neural network is a Softmax function. Since partial discharge detection is a binary classification problem, the Softmax function acts as a binary classifier, and its output can be calculated as follows.

[0052]

[0053] Among them, PD and The presence or absence of discharge is indicated in the output results, W and D are the weight matrix parameters and bias parameters, and Z is the input to the Softmax function.

[0054] The design of the multi-scale convolutional neural network layer described in step S2 mainly includes two parts: the selection of the convolutional kernel scale and the determination of the number of convolutional kernels, where the number of convolutional kernels is represented by n. Based on the sample length and the current setting of the convolutional kernel scale of the one-dimensional convolutional neural network, the convolutional kernel scale is selected from 1×2, 1×5, 1×10, 1×20, 1×32, and 1×64 to obtain features with different precisions from high frequency to low frequency.

[0055] The design concept of the two-layer multi-scale convolutional neural network is as follows. The first layer uses a convolutional kernel with a large scale range to extract features of different fineness from the training samples over a wider range. The second layer uses a convolutional kernel with a relatively small scale range to refine the previously extracted features and extract high-frequency features. Simultaneously, a comparative test was conducted on the number of convolutional kernels of different scales in the two multi-scale convolutional layers.

[0056] Based on the above design concept, the kernel sizes in multi-scale convolutional neural network layers are listed, such as... Figure 8 As shown, and nine possible combinations of two multi-scale convolutional neural network layers are given, such as... Figure 9 As shown.

[0057] The pooling mode is set to max pooling, and the pooling core size is also set accordingly. Figure 9 As shown in the diagram. To prevent overfitting, Dropout with a coefficient of 0.25 is set before the first fully connected layer, meaning that the fully connected layer is dropped with a probability of 0.25 in each training iteration. Dropout with a coefficient of 0.5 is set before the second fully connected layer. Cross-entropy loss is chosen as the loss function. Batch gradient descent is chosen as the optimizer, with a batch size of 128 and a learning rate of 0.0001. Each setup is run for 200 epochs, and the model is coded using PyTorch in Python.

[0058] In step S3, the training samples are divided into 5 layers, and the accuracy is calculated through 5-layer cross-validation. The performance metric of the classifier based on the multi-scale one-dimensional convolutional neural network is the average accuracy of the 5 training iterations. Partial discharge detection is a binary task; the classifier is considered effective only when its accuracy is greater than 50%. The accuracy of each layer is calculated using the following formula:

[0059]

[0060] The accuracy of nine settings in the three-phase signal was compared, and the results are as follows: Figure 7 As shown. To demonstrate the effectiveness of the algorithm on low-sampling-rate data, the dataset was downsampled, and voltage signals at 20MHz and 10MHz were obtained. Comparative tests were also conducted, such as... Figure 7 As shown.

[0061] from Figure 7 It can be seen that when the number of convolutional kernels in the two multi-scale convolutional layers is n1=3 and n2=3, the detection accuracy can reach 90%, and there is no significant difference in accuracy when the number of convolutional kernels increases. A sampling rate below 20MHz will affect accuracy. When the sampling rate exceeds 20MHz, the improvement in model accuracy is not significant. Furthermore, considering that increasing the network size and sampling rate will consume more computational and memory resources, leading to a decrease in training speed, setting the number of convolutional kernels in the two multi-scale convolutional neural network layers to 3, and the stride to 3 and 2 respectively, is sufficient. Setting the padding style to the same padding pattern avoids the loss of edge information and ensures that the output dimension of the parallel convolutional layers in the multi-scale convolutional neural network layers is consistent for feature concatenation.

[0062] Furthermore, the receiver operating characteristic (ROC) of the experiment was calculated by varying the probability threshold to determine whether the signal exhibited partial discharge. A key characteristic of the ROC curve is that it remains constant regardless of changes in the distribution of positive and negative samples in the test set. In the partial discharge dataset used in this paper, class imbalance occurs due to the significantly higher probability of partial discharge—negative samples (signals without partial discharge) outnumbering positive samples (signals with partial discharge). The ROC at setting 3 is shown below. Figure 10 As shown, FPR represents the false positive rate and TPR represents the true positive rate. Figure 10 This indicates that MS-Conv1D is a robust classifier for a small number of positive samples.

[0063] In summary, this invention provides a power grid fault detection model based on a multi-scale one-dimensional convolutional neural network, which solves the problem of partial discharge diagnosis in transmission lines and achieves good results under different frequency data and different convolutional kernel settings. The multi-scale one-dimensional convolutional neural network model is suitable for the diagnosis of partial discharge signals in single-phase and three-phase circuits. This model does not require specifying the length of the input sequence, nor does it require manual extraction of input signal features.

[0064] Although the present invention has been described in detail through the preferred embodiments above, it should be understood that the above description should not be considered as a limitation of the present invention. Various modifications and substitutions to the present invention will be apparent to those skilled in the art after reading the above description. Therefore, the scope of protection of the present invention should be defined by the appended claims.

Claims

1. A method of detecting partial discharge in a power transmission line, characterized by, The method comprises the following steps: S1, obtaining original power line electrical signals, and dividing the original power line electrical signals into training samples and test samples; S2, batch inputting the training samples into a multi-scale one-dimensional convolutional neural network, setting a threshold, setting a one-dimensional convolutional neural network structure and a convolution kernel size, performing forward propagation layer by layer to obtain features of the training samples, obtaining an actual output, and calculating an error between an expected output and the actual output; S3, performing backward propagation of the error from the output of the multi-scale one-dimensional convolutional neural network in S1 to the input layer by layer to update network parameters; S4, repeating steps S2 and S3 until all training samples are completed, calculating precision of the multi-scale one-dimensional convolutional neural network on the test samples, if the precision is greater than the set threshold, an ideal network is obtained, otherwise, returning to step S2; S5, outputting the precision of the ideal network on the test samples; The S2 comprises the following contents: A single sample of the original power line electrical signal has a length T and a phase p, and is represented by a column vector as follows: φ p [1:T] = [φ p,1 ; φ p,2 ;..., φ p,T ] ; The max-pooling layer performs subsampling with a step size of d, the input of the max-pooling layer is X, the output is Mp(X), the tth element of Mp(X) is t (X) = max(X d(t-1)+i , 1≤i≤a), where a is the pooling size; The input of a one-dimensional convolutional neural network layer with a Relu activation function is Y, and the output is denoted as Conv(Y), the tth element of Conv(Y) is where c is the size of the convolution kernel, ω i is the i th weight of the convolution kernel, b t is the bias of the tth element of the output of the convolution layer; The Relu activation function is: Relu(x) = max(0, x), and the function runs according to the rule of taking the maximum value between 0 and x; The flattening layer connects three-phase signals into a one-dimensional vector by using the Relu activation function; each neuron in the fully connected layer is fully connected with all neurons in the previous layer to integrate local information according to different categories in the convolutional neural network layer or the pooling layer.

2. The power transmission line partial discharge detection method of claim 1, wherein, The original power line electrical signal comprises voltage measurement signals of three-phase overhead power lines, and the measurement of each phase voltage measurement signal is performed simultaneously.

3. The power transmission line partial discharge detection method of claim 1, wherein, The multi-scale one-dimensional convolutional neural network comprises two multi-scale one-dimensional convolutional neural network layers and two fully connected layers, and finally outputs a result through a Softmax function.

4. The power transmission line partial discharge detection method of claim 3, wherein, Each one-dimensional convolutional neural network layer uses different scale convolution kernels to perform convolution operation on different scale input signals to realize extraction of different scale signal features.

5. The power transmission line partial discharge detection method of claim 1, wherein, The last layer of the multi-scale one-dimensional convolutional neural network is a Softmax function; since the partial discharge detection is a binary classification problem, the Softmax function is a binary classifier at this time, and the output can be calculated as follows, where PDand where PDand where PDand 6. The power transmission line partial discharge detection method of claim 3, wherein, If the input signal is a three-phase waveform, the output of the multi-scale convolutional neural network layer needs to be connected to a flattening layer to adjust the vector form of the fully connected layer.

Images