A fault diagnosis method for AC / DC hybrid power grid based on RPM and T-ShuffleNet V2 neural network

Abstract

The present invention discloses a method for diagnosing AC / DC hybrid power grid faults based on RPM and T‑ShuffleNet V2 neural network, the steps of which include: forming a fault voltage information matrix and a fault current information matrix; filtering the fault voltage information matrix and the fault current information matrix; expanding the fault voltage information matrix and the fault current information matrix into one column by column, respectively, and converting them into a fault voltage two-dimensional feature map and a fault current two-dimensional feature map; using the trained T‑ShuffeNet V2 neural network model to perform fault identification on the fault voltage two-dimensional feature map and the fault current two-dimensional feature map, and obtaining the fault type of each line of the AC / DC hybrid power grid to be diagnosed. Experimental results show that the method proposed by the present invention improves the accuracy and convergence speed, is less affected by transition resistance, fault line, and fault distance, and has strong anti-noise and anti-loss capabilities.

Description

Technical Field

[0001] The present invention relates to a power grid fault diagnosis method, in particular to an AC / DC hybrid power grid fault diagnosis method based on RPM and T-ShuffleNet V2 neural network. Background Art

[0002] Due to geographical conditions, natural environment and other factors, my country's energy and load space are uneven, and there is a large-scale demand for long-distance transmission of new energy. In terms of long-distance transmission, high-voltage direct current transmission has the advantages of low line cost, small transmission loss and large transmission capacity compared with traditional alternating current transmission. Under the "dual carbon" goal, in order to further solve the problem of regional imbalance of load and energy, high-voltage direct current transmission has been widely used in power transmission, and AC / DC hybrid power grid has become an inevitable trend. The mutual influence between AC and DC during the operation of AC / DC hybrid power grid can easily cause large power grid security incidents and cause serious fault losses. Therefore, timely analysis of the fault characteristics of AC / DC hybrid power grids and determination of fault types are crucial to the stability of power grid operation.

[0003] Due to the mutual influence of transient processes of AC and DC transmission lines, it is difficult to analyze the characteristics of AC / DC hybrid power grids, and traditional power grid fault analysis methods are no longer applicable. Existing studies have shown that the protection scheme of traditional AC transmission lines is no longer applicable to AC transmission lines on the inverter side of AC / DC hybrid power grids. Therefore, a new criterion for the radius of curvature of the first-traveling wave is proposed. This criterion can distinguish between A-phase short-circuit grounding faults within the area and three-phase short-circuit faults outside the area, and has strong tolerance to transition resistance and anti-noise interference. Existing studies have shown that criterion-based diagnostic algorithms can effectively diagnose faults within and outside the area when the threshold is set reasonably, but this type of algorithm often improves a single problem and has limitations in large AC / DC hybrid power grids.

[0004] With the development of artificial intelligence technology, new ideas have been provided for fault diagnosis of large AC / DC hybrid power grids. This type of algorithm has a strong learning ability. It can be used to learn the characteristics of various types of faults in each line to solve the fault diagnosis problem of AC / DC hybrid power grids. In the prior art, by collecting wide-area information of AC / DC hybrid power grids, the convolutional neural network (CNN) is trained to realize the fault line identification and fault type identification of AC / DC hybrid power grids, but the method has weak anti-noise ability and considers that the maximum transition resistance is 50Ω. In the prior art, there is also fault identification for the DC line of AC / DC hybrid power grids, which uses Fourier transform to extract the high-frequency quantity of fault current and phase mode transform to extract the voltage change to form fault characteristics, and is classified through deep belief network (DBN), realizing the fault area judgment and fault pole selection of the DC line of AC / DC hybrid power grid. This method is not easily affected by transition resistance, but has poor anti-noise ability.

[0005] In the field of artificial intelligence, algorithms have greater advantages in processing two-dimensional images than one-dimensional signals. In order to make better use of this advantage, one-dimensional signals can be converted into two-dimensional images and classified by artificial intelligence algorithms. This method has also been widely used in the field of fault diagnosis. In the prior art, transient fault voltage time domain data of the flexible DC power grid is collected, then converted into a two-dimensional image through the Gramian Angular Field (GAF), and then the Transformer neural network is improved for classification and identification to achieve fault diagnosis of the DC transmission system.

[0006] In summary, the existing technology has achieved good results in fault diagnosis of AC / DC hybrid power grids, but its ability to deal with noise interference has not yet achieved ideal results, and its ability to resist noise interference needs to be improved. At the same time, it has not fully considered the impact of data loss during signal transmission. Summary of the invention

[0007] The purpose of the invention is to provide an AC / DC hybrid power grid fault diagnosis method based on RPM and T-ShuffleNet V2 neural network, which has good anti-interference ability and can reduce the impact of data loss during signal transmission.

[0008] Technical solution: The AC / DC hybrid power grid fault diagnosis method based on RPM and T-ShuffleNet V2 neural network described in the present invention comprises the following steps:

[0009] Step 1, extracting the fault information of each line of the AC / DC hybrid power grid to be diagnosed, and forming a fault voltage information matrix and a fault current information matrix;

[0010] Step 2, using a wavelet threshold filtering algorithm to filter each column of voltage signal and current signal in the fault voltage information matrix and the fault current information matrix;

[0011] Step 3, expand the fault voltage information matrix and the fault current information matrix after filtering into one column by column, obtain one column of fault voltage information and one column of fault current information, and then use the relative position matrix to convert one column of fault voltage information and one column of fault current information into a two-dimensional fault voltage feature map and a two-dimensional fault current feature map respectively;

[0012] Step 4: Use the trained T-ShuffeNet V2 neural network model to perform fault identification on the fault voltage two-dimensional feature map and the fault current two-dimensional feature map to obtain the fault type of each line of the AC / DC hybrid power grid to be diagnosed.

[0013] Furthermore, in step 1, the specific steps of forming the fault voltage information matrix and the fault current information matrix include:

[0014] Step 1.1: Collect the data of the AC / DC hybrid power grid to be diagnosed. i The voltage signal at both ends of the DC line and j The voltage signals of the AC lines are obtained, and the obtained voltage signals are sorted by columns to form a fault voltage information matrix:

[0015] G U =[ U 1a , U 1b , U 2a , U 2b , … , U ia , U ib , U i+1 , U i+2 , … , U i+j ], where U ia and U ib Indicates i The voltage signal at both ends of the DC line, U i+j Indicates j The voltage signal of an AC line;

[0016] Step 1.2: Collect the data from the AC / DC hybrid power grid to be diagnosed. i The current signal at both ends of the DC line and j The current signals of the AC lines are obtained, and the obtained current signals are sorted by columns to form the fault current information matrix:

[0017] G I =[ I 1a , I 1b , I 2a , I 2b , … , I ia , I ib , Ii+1 , I i+2 , … , I i+j ], where I ia and I ib Indicates i The current signal at both ends of the DC line, I i+j Indicates j The current signal of an AC line;

[0018] Step 1.3, using a sliding time window to perform data enhancement on each column of voltage signal and current signal in the fault voltage information matrix and the fault current information matrix.

[0019] Furthermore, in step 1.3, the window width of the sliding window is 20 ms and the step length is 0.1 ms.

[0020] Furthermore, in step 3, the specific steps of using the relative position matrix to convert a column of fault voltage information and a column of fault current information into a two-dimensional fault voltage feature map and a two-dimensional fault current feature map are as follows:

[0021] Step 3.1: Set a column of fault voltage information and a column of fault current information as independent time series X , X ={ x 1 , x 2 ,… x t ,…, x n},in x t Represents the information value at the corresponding moment, and then performs Z-score standardization to obtain the standard normal distribution:

[0022] Z ={ z 1 , z 2 ,… z t ,…, z n}, where z t =( x t - μ ) / σ , t =1,2,…, n , μyes X The average value of σ yes X The standard deviation of

[0023] Step 3.2, then use the piecewise aggregation approximation method to reduce the numerator k , thus changing the dimension from n Down to m , and generate a new time series: , where For time series The i element, ,when Sometimes, there is ,when Sometimes, there is , where is the ceiling operator, is the floor operator;

[0024] Step 3.3, calculate different moments The relative position of The matrix of relative position relationship between M for: ;

[0025] Step 3.4, use min - max The normalization algorithm transforms the matrix M Convert to gray value matrix and get relative position matrix F for: , where min ( M ) represents the matrix M The minimum value in max ( M ) represents the matrix M The maximum value in the relative position matrix F That is, the two-dimensional characteristic diagram of fault voltage and the two-dimensional characteristic diagram of fault current.

[0026] Furthermore, in step 4, the T-ShuffeNet V2 neural network model includes sequentially connected Conv1 layer, MaxPool layer, Stage2 layer, Stage3 layer, Stage4 layer, Conv5 layer, GAP layer and FC layer; the Conv1 layer is a 3×3 convolution layer; the Stage2 layer and the Stage4 layer are each composed of a downsampling module and three basic modules; the Stage3 layer is composed of a downsampling module and seven basic modules; the Conv5 layer is a 1×1 convolution layer.

[0027] Furthermore, the downsampling module includes a deformable convolution layer, a left convolution branch, a right convolution branch, a Concat layer and a Channel Shuffle layer; the left convolution branch includes a 3×3 DWConv layer and a 1×1 Conv layer connected sequentially; the right convolution branch includes a 1×1 Conv layer, a 3×3 DWConv layer and a 1×1 Conv layer connected sequentially; the output features of the deformable convolution layer are simultaneously sent to the left convolution branch and the right convolution branch for corresponding convolution processing; the Concat layer is used to connect the output features of the left convolution branch and the right convolution branch; the Channel Shuffle layer is used to perform channel information fusion on the output features of the Concat layer.

[0028] Furthermore, the basic module includes a Channel split layer, a left identity branch, a right multi-convolution branch, a Concat layer and a Pixel-Channel Shuffle layer; the left identity branch is used for identity mapping; the right multi-convolution branch includes a sequentially connected 1×1 Conv layer, a multi-branch strip convolution layer and a 1×1 Conv layer; the Channel split layer is used to divide the input features into channels, and after division, they are respectively sent to the left identity branch and the right multi-convolution branch; the Concat layer is used to connect the output features of the left identity branch and the right multi-convolution branch; the Pixel-ChannelShuffle layer is used to perform dual fusion of pixels and channels on the output features of the Concat layer.

[0029] Further, the Pixel-Channel Shuffle layer includes a Pixel Shuffle layer and a Channel Shuffle layer connected sequentially;

[0030] The Pixel Shuffle layer is used to fuse the input features by pixels; the Channel Shuffle layer is used to recombine the channels.

[0031] Furthermore, the Pixel Shuffle layer includes two parallel fusion branches, which include a sequentially connected Pixel spli layer, a 3×3 Conv layer, and a Pixel merge layer; the Channel Shuffle layer is a 1×1 convolution layer.

[0032] Furthermore, the multi-branch strip convolution layer includes a 1×1 Conv layer, three convolution branches, a Concat layer, a 1×1 Conv layer, and an Add layer; the first convolution branch is composed of a 1×3 strip convolution, a 3×1 strip convolution, and a dilated convolution with a dilation rate of 1 and a size of 3×3, which are sequentially connected; the second convolution branch is composed of a 1×5 strip convolution, a 5×1 strip convolution, and a dilated convolution with a dilation rate of 2 and a size of 3×3, which are sequentially connected; the third convolution branch is composed of a 1×7 strip convolution, which is sequentially connected , 7×1 strip convolution, and a 3×3 dilated convolution with a dilation rate of 3. The input features are first convolved through a 1×1 Conv layer to reduce the number of parameters, and then three sets of features are obtained through three convolution branches. These three sets of features are then fused through a Concat layer. The fused features are then convolved through a 1×1 Conv layer to restore them to the input size. Finally, an Add layer is used to perform an Add operation to fuse the features restored to the input size with the original features.

[0033] Compared with the prior art, the present invention has the following beneficial effects: (1) The method proposed in the present invention collects the voltage and current signals of the power grid, processes the voltage and current signals respectively through wavelet threshold filtering and RPM, and performs pixel splicing to form a two-dimensional feature map. The two-dimensional feature map can well represent the time series information and improve the fault type recognition accuracy of the AC / DC hybrid power grid; (2) The Pixel-Channel Shuffle structure proposed in the present invention integrates global information, combines multi-branch structure, dilated convolution and strip convolution, and replaces the deep separable convolution structure of the basic module in the original network with a multi-branch strip convolution structure; the receptive field of the model is increased by using deformable convolution; experimental results show that the proposed method improves the accuracy and convergence speed, and is less affected by transition resistance, fault line and fault distance, and has strong anti-noise and anti-loss capabilities; (3) Compared with some current deep learning algorithms Alexnet, MobileNet V2 and ResNet, the T-ShuffeNet V2 neural network proposed in the present invention shows excellent performance in extracting deep feature information of the RPM feature map and can accurately identify the fault type of the AC / DC hybrid power grid. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 is a flow chart of the method of the present invention;

[0035] Figure 2 It is a schematic diagram of the relative position matrix conversion process of the present invention;

[0036] Figure 3 Schematic diagram of the T-ShuffleNet V2 neural network model of the present invention;

[0037] Figure 4 It is a schematic diagram of the basic modules of the present invention;

[0038] Figure 5 It is a schematic diagram of a downsampling module of the present invention;

[0039] Figure 6 It is a schematic diagram of the Pixel-Channel Shuffle layer of the present invention;

[0040] Figure 7 Schematic diagram of a multi-branch strip convolutional layer of the present invention;

[0041] Figure 8 Schematic diagram of a deformable convolutional layer of the present invention;

[0042] Fig. 9 A power grid fault simulation model diagram of the present invention;

[0043] Fig.10 It is a sample diagram of each fault type of the present invention;

[0044] Fig.11 It is the Loss change curve diagram of the present invention;

[0045] Fig.12 This is a test set accuracy change curve diagram of the present invention. DETAILED DESCRIPTION

[0046] The technical solution of the present invention is described in detail below with reference to the accompanying drawings, but the protection scope of the present invention is not limited to the embodiments.

[0047] like Figure 1 As shown, the AC / DC hybrid power grid fault diagnosis method based on RPM and T-ShuffleNet V2 neural network disclosed in the present invention includes the following steps:

[0048] Step 1, extracting the fault information of each line of the AC / DC hybrid power grid to be diagnosed, and forming a fault voltage information matrix and a fault current information matrix;

[0049] Step 2, using a wavelet threshold filtering algorithm to filter each column of voltage signal and current signal in the fault voltage information matrix and the fault current information matrix;

[0050] Step 3, expand the fault voltage information matrix and the fault current information matrix after filtering into one column by column, obtain one column of fault voltage information and one column of fault current information, and then use the relative position matrix to convert one column of fault voltage information and one column of fault current information into a two-dimensional fault voltage feature map and a two-dimensional fault current feature map respectively;

[0051] Step 4: Use the trained T-ShuffeNet V2 neural network model to perform fault identification on the fault voltage two-dimensional feature map and the fault current two-dimensional feature map to obtain the fault type of each line of the AC / DC hybrid power grid to be diagnosed.

[0052] Furthermore, the wavelet threshold filtering algorithm mainly consists of three parts: wavelet decomposition, wavelet decomposition coefficient threshold quantization and signal reconstruction. First, the wavelet basis function is used to perform multi-scale wavelet decomposition on the data to obtain low-frequency signals and high-frequency signals; then the wavelet threshold function is used to process the high-frequency signal to determine and remove signal noise; finally, the low-frequency signal and the high-frequency signal after threshold processing are reconstructed through wavelet inverse transform to obtain the denoised signal.

[0053] Furthermore, in step 1, the specific steps of forming the fault voltage information matrix and the fault current information matrix include:

[0054] Step 1.1: Collect the data of the AC / DC hybrid power grid to be diagnosed. i The voltage signal at both ends of the DC line and j The voltage signals of the AC lines are obtained, and the obtained voltage signals are sorted by columns to form a fault voltage information matrix:

[0055] G U =[ U 1a , U 1b , U 2a , U 2b , … , U ia , U ib , U i+1 , U i+2 , … , U i+j ], where U ia and U ib Indicates i The voltage signal at both ends of the DC line, U i+j Indicates j The voltage signal of an AC line;

[0056] Step 1.2: Collect the data from the AC / DC hybrid power grid to be diagnosed. i The current signal at both ends of the DC line and jThe current signals of the AC lines are obtained, and the obtained current signals are sorted by columns to form the fault current information matrix:

[0057] G I =[ I 1a , I 1b , I 2a , I 2b , … , I ia , I ib , I i+1 , I i+2 , … , I i+j ], where I ia and I ib Indicates i The current signal at both ends of the DC line, I i+j Indicates j The current signal of an AC line;

[0058] Step 1.3, using a sliding time window to perform data enhancement on each column of voltage signal and current signal in the fault voltage information matrix and the fault current information matrix.

[0059] Furthermore, in step 1.3, the window width of the sliding window is 20 ms and the step length is 0.1 ms.

[0060] Further, such as Figure 2 As shown, in step 3, the specific steps of using the relative position matrix to convert a column of fault voltage information and a column of fault current information into a two-dimensional fault voltage feature map and a two-dimensional fault current feature map are as follows:

[0061] Step 3.1: Set a column of fault voltage information and a column of fault current information as independent time series X , X ={ x 1 , x 2 ,… x t ,…, x n},in x tRepresents the information value at the corresponding moment, and then performs Z-score standardization to obtain the standard normal distribution:

[0062] Z ={ z 1 , z 2 ,… z t ,…, z n}, where z t =( x t - μ ) / σ , t =1,2,…, n , μ yes X The average value of σ yes X The standard deviation of

[0063] Step 3.2, then use the piecewise aggregation approximation method to reduce the numerator k , thus changing the dimension from n Down to m , and generate a new time series: , where For time series The i element, ,when Sometimes, there is ,when Sometimes, there is , where is the ceiling operator, is the floor operator;

[0064] Step 3.3, calculate the relative positions at different times and obtain a matrix used to characterize the relative position relationship between different time series M for: ;

[0065] Step 3.4, use min - max The normalization algorithm transforms the matrix M Convert to gray value matrix and get relative position matrix F for: , where min ( M ) represents the matrix M The minimum value in max ( M ) represents the matrix MThe maximum value in the relative position matrix F That is, the two-dimensional characteristic diagram of fault voltage and the two-dimensional characteristic diagram of fault current.

[0066] Furthermore, in step 4, if Figure 3 As shown in the figure, the T-ShuffeNet V2 neural network model includes a sequentially connected Conv1 layer, MaxPool layer, Stage2 layer, Stage3 layer, Stage4 layer, Conv5 layer, GAP layer and FC layer; the Conv1 layer is a 3×3 convolution layer; the Stage2 layer and the Stage4 layer are each composed of a downsampling module and three basic modules; the Stage3 layer is composed of a downsampling module and seven basic modules; the Conv5 layer is a 1×1 convolution layer.

[0067] Further, such as Figure 5 As shown in the figure, the downsampling module includes a deformable convolution layer, a left convolution branch, a right convolution branch, a Concat layer and a Channel Shuffle layer; the left convolution branch includes a 3×3 DWConv layer and a 1×1 Conv layer connected sequentially; the right convolution branch includes a 1×1 Conv layer, a 3×3 DWConv layer and a 1×1 Conv layer connected sequentially; the output features of the deformable convolution layer are simultaneously sent to the left convolution branch and the right convolution branch for corresponding convolution processing; the Concat layer is used to connect the output features of the left convolution branch and the right convolution branch; the ChannelShuffle layer is used to perform channel information fusion on the output features of the Concat layer.

[0068] Deformable convolutional layers such as Figure 8 As shown in the figure, by setting the deformable convolution layer, the receptive field can be enhanced, the overall characteristics of the object can be learned, and the sampling position is more consistent with the shape and size of the object itself. The convolution formula of the deformable convolution layer is: ,

[0069] In the formula, p 0 is any point on the input feature map, p n Represents the offset of each point in the convolution kernel compared to the center point, w ( p n ) is the weight of the corresponding position, R represents the sampling area of ​​the standard convolution kernel, x ( p 0 + p n + Δ p n) is the input feature map p 0 + p n + Δ p n The value at position, y ( p 0 ) is the convolution result, Δ p n The offset introduced for each point, offset Δ p n It is generated by performing an additional convolution operation on the input feature map.

[0070] Further, such as Figure 4 As shown in the figure, the basic module includes a Channel split layer, a left identity branch, a right multi-convolution branch, a Concat layer and a Pixel-Channel Shuffle layer; the left identity branch is used for identity mapping; the right multi-convolution branch includes a sequentially connected 1×1 Conv layer, a multi-branch strip convolution layer and a 1×1 Conv layer; the Channelsplit layer is used to divide the input features into channels, and after division, they are respectively sent to the left identity branch and the right multi-convolution branch; the Concat layer is used to connect the output features of the left identity branch and the right multi-convolution branch; the Pixel-Channel Shuffle layer is used to perform dual fusion of pixels and channels on the output features of the Concat layer.

[0071] Further, such as Figure 6 As shown, the Pixel-Channel Shuffle layer includes a sequentially connected Pixel Shuffle layer and a Channel Shuffle layer; the Pixel-Channel Shuffle layer is used for shuffling pixels and channels; the PixelShuffle layer is used for pixel feature fusion of input features; and the Channel Shuffle layer is used for recombining channels.

[0072] Furthermore, the Pixel Shuffle layer includes two parallel fusion branches, the fusion branch includes a sequentially connected Pixel spli layer, a 3×3 Conv layer and a Pixel merge layer; the Channel Shuffle layer is a 1×1 convolution layer; the Pixel spli layer is used for pixel division; and the Pixel merge layer is used for pixel merging.

[0073] The Pixel-Channel Shuffle layer can help the network grasp the global information, realize the pixel fusion and channel fusion of features, enhance the feature extraction capability of the T-ShuffleNet V2 neural network, and is suitable for fault type identification in AC / DC hybrid power grids.

[0074] Further, such as Figure 7 As shown in the figure, the multi-branch strip convolution layer includes a 1×1 Conv layer, three convolution branches, a Concat layer, a 1×1 Conv layer and an Add layer; the first convolution branch consists of a sequentially connected 1×3 strip convolution, a 3×1 strip convolution and a dilated convolution with a dilation rate of 1 and a size of 3×3; the second convolution branch consists of a sequentially connected 1×5 strip convolution, a 5×1 strip convolution and a dilated convolution with a dilation rate of 2 and a size of 3×3; the third convolution branch consists of a sequentially connected 1×7 strip convolution, The input feature is first convolved through a 1×1 Conv layer to reduce the number of parameters, and then three sets of features are obtained through three convolution branches. These three sets of features are then fused through a Concat layer, and the fused features are then convolved through a 1×1 Conv layer to restore them to the input size. Finally, an Add layer is used to perform an Add operation to fuse the features restored to the input size with the original features.

[0075] By setting strip convolutions of different sizes and dilated convolutions of different dilation rates in the multi-branch strip convolution layer, the neural network pays more attention to long-distance information, improves the accuracy of fault type recognition in AC / DC hybrid power grids, and enhances the network's ability to extract time series information.

[0076] In order to verify the reliability of the AC / DC hybrid power grid fault diagnosis method based on RPM and T-ShuffleNet V2 neural network of the present invention, a simulation experiment was carried out as follows.

[0077] In PSCAD / EMTDC software, Fig. 9 The power grid fault simulation model shown in the figure, where G1 is a doubly fed wind turbine generator set, G2, G3, G4, and G5 are thermal generator sets, and a 12-pulse thyristor converter is used as a DC line converter. This experiment built an environment and conducted training on a deep learning workstation. The workstation hardware configuration is CPU (lntel Core i9-12900H), GPU (NVIDIV RTX A2000 Laptop-8G); the software environment is Pycharm; the programming language is Python3.7, and the Pytorch deep learning framework. The power grid parameters of the power grid fault simulation model are shown in Table 1.

[0078] Table 1 AC / DC hybrid power grid parameters

[0079]

[0080] The sample data of the present invention comes from Fig. 9 The AC / DC power grid simulation model shown. Since the faults occurring in the DC lines are mainly single-stage grounding faults, for the two DC lines, the present invention sets up single-stage grounding fault simulations with different fault distances, and the transition resistance setting range is 0.1Ω-200Ω. For the 10 AC lines, the present invention sets up 10 fault types, including ABC three-phase short circuit, AB two-phase short circuit, AB two-phase ground short circuit, AC two-phase short circuit, AC two-phase ground short circuit, BC two-phase short circuit, BC two-phase ground short circuit, A ground short circuit, B ground short circuit, and C ground short circuit, and the transition resistance setting range is 0.1Ω-200Ω. The data in the time window of 5ms before the fault and 16ms after the fault is collected at a sampling frequency of 10kHz, and the data is enhanced through a sliding time window, and then processed using wavelet threshold filtering and converted into a two-dimensional image through RPM. The samples of each fault type are shown as follows. Fig.10 A total of 11,000 fault samples are obtained for the 11 types of faults, and the composition of the data set is shown in Table 2.

[0081] Table 2 Dataset composition table

[0082]

[0083] The size of the line fault feature map is set to 800×400, and the data set is divided into a training set and a test set at a ratio of 8:2. The fault type is identified by the T-ShuffeNet V2 neural network model, and the network epochs is set to 100, the batch size is set to 16, and the learning rate is set to 0.0001. The loss value of the T-ShuffeNet V2 neural network and the accuracy change curve of the test set are shown in the figure. Fig.11 , Fig.12 As shown in the figure, it can be seen that after each new improvement point is added, the convergence speed of the algorithm is improved, and finally the fault type recognition accuracy rate reaches 100%. The recognition accuracy rates of various fault types are shown in Table 3. It can be seen that the method proposed in the present invention is effective for the fault type recognition problem of AC / DC hybrid power grid, and can effectively distinguish 11 types of faults under different transition resistances, with excellent performance.

[0084] Table 3. Accuracy of T-ShuffeNet V2 in identifying various fault types

[0085]

[0086] In order to verify that this image encoding method has advantages in identifying the fault type of AC / DC hybrid power grid, the relative position matrix (RPM) is compared with the Gram-Angle Difference Field (GADF), Gram-Angle Summation Fields (GASF), Markov Transition Field (MTF), and Short-Time Fourier Transform (STFT) in an experiment. The ShuffleNet V2 neural network is used to identify the fault type of AC / DC hybrid power grid. The results of the comparative experiment are shown in the following table:

[0087] Table 4 Comparison of fault type recognition results of different image coding methods

[0088]

[0089] It can be seen from the above table that in the AC / DC hybrid power grid fault recognition task, the ShuffleNet V2 neural network has difficulty extracting feature information from the STFT image, resulting in a low recognition rate. Compared with MTF, GADF, and GASF, the image coding method RPM used in the present invention has better effect.

[0090] There is noise interference in the actual power grid operation, especially the fault information transmission process is easily interfered by surrounding electronic devices, forming normally distributed noise, which is similar to Gaussian white noise. Therefore, the present invention forms fault type samples by adding 15dB, 20 dB, 25dB, and 30dB Gaussian white noise to the fault information matrix, and then forms fault type samples through wavelet threshold filtering and RPM processing. The number of samples of each fault type is 100, for a total of 1100 samples, and the fault line is identified by the trained T-ShuffeNet V2 neural network. At the same time, in order to verify the effectiveness of the improvement points of the present invention, the same noise interference experiment was carried out on the ShuffeNet V2 neural network under different improvement points, and the results are shown in Tables 5 and 6:

[0091] Table 5 Fault type recognition accuracy under different noise interference

[0092]

[0093] Table 6 Accuracy of T-ShuffeNet V2 in identifying various fault types under 15dB noise interference

[0094]

[0095] From the results in the above table, it can be seen that the accuracy of the original ShuffeNet V2 neural network drops significantly when facing noise, but after adding Pixel-Channel Shuffle, it can cope with 20dB noise interference. The addition of multi-branch strip convolution improves the accuracy of the model by 8.7% under the interference of 15dB noise. Finally, deformable convolution is added to improve the convergence speed and slightly improve the noise resistance of the algorithm. Under 15dB noise interference, T-ShuffeNet V2 has poor recognition effect on B-phase ground short circuit and AB two-phase ground short circuit faults, but still has an accuracy of 89% and 89.9%; the recognition accuracy of ABC three-phase short circuit, AB two-phase short circuit, and AC two-phase short circuit faults is as high as more than 99%. The experimental results show that T-ShuffeNet V2 has strong robustness to noise and can still accurately identify the fault type under the interference of low signal-to-noise ratio noise.

[0096] Since the AC / DC hybrid power grid is susceptible to interference and may cause data loss during the data collection process. Therefore, in order to verify the ability of the algorithm of the present invention to resist data loss, after performing 5%, 10%, 15%, and 20% random data loss processing on the fault line signal, fault type samples are formed through wavelet threshold filtering and RPM processing. The number of samples of each fault type is 100, for a total of 1100 samples, and the fault line is identified by the trained T-ShuffeNet V2 neural network. At the same time, in order to verify the effectiveness of the improvement points of the present invention, the same loss interference experiment is carried out on the ShuffeNet V2 neural network under different improvement points, and the results are shown in Tables 7 and 8 below:

[0097] Table 7 Fault type recognition accuracy under different data loss conditions

[0098]

[0099] Table 8 Accuracy of T-ShuffeNet V2 in identifying various fault types under 20% data loss interference

[0100]

[0101] As can be seen from the table, when the fault line has signal loss, the original ShuffeNet V2 neural network has poor fault type recognition ability. After adding Pixel-Channel Shuffle, the accuracy increased by about 4% under each loss ratio. After adding multi-branch strip convolution, the accuracy increased by about 2% under each loss ratio. When the fault line data is lost by 20%, there are 9 types of fault types with an identification accuracy greater than 94%, and 4 types of fault types with an identification accuracy greater than 98%. The experimental results show that T-ShuffeNet V2 has a high accuracy rate when facing different data loss interferences, can accurately identify fault types, and has high reliability.

[0102] In order to verify that the network model of the present invention has advantages in identifying the fault type of AC / DC hybrid power grid, the T-ShuffleNet V2 neural network is compared with the nearest neighbor classification algorithm (K-Nearest Neighbor, KNN), Alexnet neural network, MobileNet V2 neural network, and ResNet34 neural network. The results of the comparative experiments are shown in the following table:

[0103] Table 9 Comparison of fault type recognition results of different neural networks

[0104]

[0105] As can be seen from the above table, the traditional machine learning algorithm KNN has difficulty in extracting effective feature information and has a low recognition accuracy when facing multi-classification fault type identification of AC / DC hybrid power grids. Although the Alexnet neural network and MobileNet V2 neural network perform well in other tasks, their accuracy can only reach 88.9% and 91% in the recognition of fault types in AC / DC hybrid power grids. The ResNet34 neural network has the characteristics of residual connection and deep trainability, and has achieved an accuracy of 98.1%, indicating that this network is good at processing complex spatial information. The improved method proposed in the present invention has achieved an accuracy of 100%, and has shown better results in the recognition of fault types in AC / DC hybrid power grids.

[0106] Considering the problem that the signal transmission process of AC / DC hybrid power grid is easily interfered, this paper proposes a fault type diagnosis method for AC / DC hybrid power grid based on RPM and T-ShuffleNet V2 neural network, and obtains the following conclusions through simulation experiments using PSCAD:

[0107] 1) The method proposed in the present invention collects the voltage and current signals of the power grid, processes the voltage and current signals respectively through wavelet threshold filtering and RPM, and performs pixel splicing to form a two-dimensional feature map. The two-dimensional feature map can well represent the time series information and improve the fault type recognition accuracy of the AC / DC hybrid power grid;

[0108] 2) The present invention proposes a Pixel-Channel Shuffle structure to integrate global information; uses a combination of multi-branch structure, dilated convolution, and strip convolution to replace the depth-separable convolution structure of the basic module in the original network with a multi-branch strip convolution structure; uses deformable convolution to increase the receptive field of the model; experimental results show that the proposed method improves the accuracy and convergence speed, is less affected by transition resistance, fault line, and fault distance, and has strong anti-noise and anti-loss capabilities;

[0109] 3) Compared with some current deep learning algorithms such as Alexnet neural network, MobileNet V2 neural network, and ResNet34 neural network, the T-ShuffeNet V2 neural network proposed in the present invention shows excellent performance in extracting deep feature information of RPM feature graphs, and can accurately identify the fault type of AC / DC hybrid power grid.

[0110] As described above, although the present invention has been shown and described with reference to specific preferred embodiments, it should not be construed as limiting the present invention itself. Various changes in form and details may be made without departing from the spirit and scope of the present invention as defined in the appended claims.

Claims

1. A fault diagnosis method for AC / DC hybrid power grid based on RPM and T-ShuffleNet V2 neural network, characterized in that: The steps include: Step 1, extracting the fault information of each line of the AC / DC hybrid power grid to be diagnosed, and forming a fault voltage information matrix and a fault current information matrix; Step 2, using a wavelet threshold filtering algorithm to filter each column of voltage signal and current signal in the fault voltage information matrix and the fault current information matrix; Step 3, expand the fault voltage information matrix and the fault current information matrix after filtering into one column by column, obtain one column of fault voltage information and one column of fault current information, and then use the relative position matrix to convert one column of fault voltage information and one column of fault current information into a two-dimensional fault voltage feature map and a two-dimensional fault current feature map respectively; Step 4: Use the trained T-ShuffeNet V2 neural network model to perform fault identification on the fault voltage two-dimensional feature map and the fault current two-dimensional feature map to obtain the fault type of each line of the AC / DC hybrid power grid to be diagnosed; In step 4, the T-ShuffeNet V2 neural network model includes a sequentially connected Conv1 layer, a MaxPool layer, a Stage2 layer, a Stage3 layer, a Stage4 layer, a Conv5 layer, a GAP layer, and a FC layer; the Conv1 layer is a 3×3 convolution layer; the Stage2 layer and the Stage4 layer are each composed of a downsampling module and three basic modules; the Stage3 layer is composed of a downsampling module and seven basic modules; the Conv5 layer is a 1×1 convolution layer; The basic module includes Channel split layer, left identity branch, right multi-convolution branch, Concat layer and Pixel-Channel Shuffle layer; the left identity branch is used for identity mapping; the right multi-convolution branch includes a sequentially connected 1×1 Conv layer, a multi-branch strip convolution layer and a 1×1 Conv layer; the Channel split layer is used to divide the input features into channels, and after division, they are sent to the left identity branch and the right multi-convolution branch respectively; the Concat layer is used to connect the output features of the left identity branch and the right multi-convolution branch; the Pixel-Channel Shuffle layer is used to perform dual fusion of pixels and channels on the output features of the Concat layer; The multi-branch strip convolution layer includes a 1×1 Conv layer, three convolution branches, a Concat layer, a 1×1 Conv layer, and an Add layer; the first convolution branch consists of a sequentially connected 1×3 strip convolution, a 3×1 strip convolution, and a dilated convolution with a dilation rate of 1 and a size of 3×3; the second convolution branch consists of a sequentially connected 1×5 strip convolution, a 5×1 strip convolution, and a dilated convolution with a dilation rate of 2 and a size of 3×3; the third convolution branch consists of a sequentially connected 1×7 strip convolution, a 7×1 strip convolution, and a dilated convolution with a dilation rate of 2 and a size of 3×3. ×1 strip convolution and a dilated convolution with a dilation rate of 3 and a size of 3×3; the input features first pass through a 1×1 Conv layer for convolution to reduce the amount of parameters, and then obtain three sets of features through three convolution branches, and then fuse these three sets of features through a Concat layer for Concat operation, and then pass the fused features through a 1×1 Conv layer for 1×1 convolution operation to restore them to the input size, and finally use the Add layer to perform the Add operation to fuse the features restored to the input size with the original features.

2. The AC / DC hybrid power grid fault diagnosis method based on RPM and T-ShuffleNet V2 neural network according to claim 1 is characterized in that: In step 1, the specific steps of forming the fault voltage information matrix and the fault current information matrix include: Step 1.1: Collect the data of the AC / DC hybrid power grid to be diagnosed. i The voltage signal at both ends of the DC line and j The voltage signals of the AC lines are obtained, and the obtained voltage signals are sorted by columns to form a fault voltage information matrix: G U =[ U 1a , U 1b , U 2a , U 2b , … , U ia , U ib , U i+1 , U i+2 , … , U i+j ], where U ia and U ib Indicates i The voltage signal at both ends of the DC line, U i+j Indicates j The voltage signal of an AC line; Step 1.2: Collect the data from the AC / DC hybrid power grid to be diagnosed. i The current signal at both ends of the DC line and j The current signals of the AC lines are obtained, and the obtained current signals are sorted by columns to form the fault current information matrix: G I =[ I 1a , I 1b , I 2a , I 2b , … , I ia , I ib , I i+1 , I i+2 , … , I i+j ], where I ia and I ib Indicates i The current signal at both ends of the DC line, I i+j Indicates j The current signal of an AC line; Step 1.3, using a sliding time window to perform data enhancement on each column of voltage signal and current signal in the fault voltage information matrix and the fault current information matrix.

3. The AC / DC hybrid power grid fault diagnosis method based on RPM and T-ShuffleNet V2 neural network according to claim 2 is characterized in that: In step 1.3, the window width of the sliding window is 20 ms and the step length is 0.1 ms.

4. The AC / DC hybrid power grid fault diagnosis method based on RPM and T-ShuffleNet V2 neural network according to claim 1, characterized in that: In step 3, the specific steps of using the relative position matrix to convert a column of fault voltage information and a column of fault current information into a two-dimensional fault voltage feature map and a two-dimensional fault current feature map are as follows: Step 3.1: Set a column of fault voltage information and a column of fault current information as independent time series X , X ={ x 1, x 2,… x t ,…, x n },in x t Represents the information value at the corresponding moment, and then performs Z-score standardization to obtain the standard normal distribution: Z ={ z 1, z 2,… z t ,…, z n }, where z t =( x t - μ ) / σ , t =1,2,…, n , μ yes X The average value of σ yes X The standard deviation of Step 3.2, then use the piecewise aggregation approximation method to reduce the numerator k , thus changing the dimension from n Down to m , and generate a new time series: , where For time series The i element, ,when Sometimes, there are ,when Sometimes, there is , where is the ceiling operator, is the floor operator; Step 3.3, calculate different moments The relative position of The matrix of relative position relationship between M for: ; Step 3.4, use min - max The normalization algorithm transforms the matrix M Convert to gray value matrix and get relative position matrix F for: , where min ( M ) represents the matrix M The minimum value in max ( M ) represents the matrix M The maximum value in the relative position matrix F That is, the two-dimensional characteristic diagram of fault voltage and the two-dimensional characteristic diagram of fault current.

5. The AC / DC hybrid power grid fault diagnosis method based on RPM and T-ShuffleNet V2 neural network according to claim 1, characterized in that: The downsampling module includes a deformable convolution layer, a left convolution branch, a right convolution branch, a Concat layer and a Channel Shuffle layer; the left convolution branch includes a 3×3 DWConv layer and a 1×1 Conv layer connected sequentially; the right convolution branch includes a 1×1 Conv layer, a 3×3 DWConv layer and a 1×1 Conv layer connected sequentially; the output features of the deformable convolution layer are simultaneously sent to the left convolution branch and the right convolution branch for corresponding convolution processing; the Concat layer is used to connect the output features of the left convolution branch and the right convolution branch; the ChannelShuffle layer is used to perform channel information fusion on the output features of the Concat layer.

6. The AC / DC hybrid power grid fault diagnosis method based on RPM and T-ShuffleNet V2 neural network according to claim 1, characterized in that: The Pixel-Channel Shuffle layer includes a sequentially connected Pixel Shuffle layer and a Channel Shuffle layer; the Pixel Shuffle layer is used to perform pixel feature fusion on the input features; and the Channel Shuffle layer is used to recombine channels.

7. The AC / DC hybrid power grid fault diagnosis method based on RPM and T-ShuffleNet V2 neural network according to claim 6, characterized in that: The Pixel Shuffle layer consists of two parallel fusion branches, which include a sequentially connected Pixel spli layer, a 3×3 Conv layer, and a Pixel merge layer; the Channel Shuffle layer is a 1×1 convolution layer.

Images