Distribution network fault distance measurement method based on phase voltage mode transformation and multi-channel data fusion

Abstract

The invention discloses a distribution network fault distance measurement method based on phase voltage mode conversion and multi-channel data fusion. The method comprises the following steps: starting a fault detection criterion by monitoring a zero-sequence voltage of a medium-voltage side; acquiring three-phase voltage data of a fault line of the power distribution network, and taking data of a period of cyclic waves as data to be processed; the three-phase voltage of the fault line is converted into a voltage line mode component through a mode transformation method; processing the voltage line mode component through an improved complementary set empirical mode decomposition algorithm to obtain an intrinsic mode function of the fault line mode component; the intrinsic mode function is converted into a feature map in an image quantization coding mode; and inputting the voltage line mode component and the feature map into a multi-channel convolutional neural network for feature mining to obtain a fault distance. According to the method, the accuracy, the real-time performance and the reliability of power distribution network fault distance measurement are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of distribution network fault detection, and in particular to a distribution network fault distance measurement method based on phase voltage mode conversion and multi-channel data fusion. Background Art

[0002] In my country's current distribution network, the neutral point non-effective grounding method is widely used, among which single-phase grounding faults account for more than 80% of the total faults. If these faults are not cleared in time, they may cause insulation breakdown and trigger multi-phase or multi-point faults, which will seriously threaten the safety and stable operation of the distribution network. With the advancement of the dual-carbon strategic goals and the increase in the proportion of new energy, the randomness and nonlinearity of the system have also increased, making the fault characteristics more complex.

[0003] Existing distribution network fault location methods mainly include matrix method, probability estimation method, artificial intelligence method, distributed measurement system, single-ended measurement system, asynchronous measurement method, generalized fault location method, sparse representation vector method and fuzzy clustering method. Most of the above methods require the comprehensive application of multiple technologies and have high computational complexity. This means that in actual operation, a large amount of computing resources and time may be required, and the real-time performance is poor, which makes it difficult to meet the needs of rapid fault response. Secondly, most methods have high requirements on the accuracy and completeness of the measurement data, and data loss or errors may seriously affect the positioning accuracy. Some methods have poor adaptability to different network configurations or load changes and cannot operate stably in various environments.

[0004] Deep learning can automatically learn potential feature patterns from complex data, thereby achieving more accurate fault diagnosis and location. In existing research, one-dimensional convolutional neural networks are widely used in fault location tasks, extracting local features in time series data through convolutional layers. However, when processing complex time series signals, one-dimensional convolutional neural networks may not be able to fully utilize the global time series information in the signal. For those fault signals with significant and varied time series features, simple one-dimensional convolution may not capture enough contextual information, affecting the accuracy of fault location. Some methods convert one-dimensional time domain signals into two-dimensional time-frequency feature maps for processing. Although they can better capture frequency domain features, the conversion process is often time-consuming and may result in the global features of the original signal not being fully utilized.

[0005] Improving the accuracy, real-time performance and reliability of fault location in distribution networks has become an urgent issue that needs to be addressed. Summary of the invention

[0006] The present invention proposes a distribution network fault location method based on phase voltage mode conversion and multi-channel data fusion, the purpose of which is to overcome the defects of the prior art and improve the accuracy, real-time performance and reliability of distribution network fault location.

[0007] The technical solution of the present invention is as follows:

[0008] A distribution network fault location method based on phase voltage mode conversion and multi-channel data fusion includes the following steps:

[0009] S1: Determine whether the zero-sequence voltage anomaly exceeds a set threshold. If so, it indicates that a ground fault has been detected in the distribution network, and step S2 is executed. Otherwise, this step is repeated.

[0010] S2: Collect the three-phase voltage data of the fault line of the distribution network, and take a period of cycle data as the data to be processed;

[0011] S3: converting the three-phase voltage of the fault line into the line mode component of the voltage signal by means of a mode conversion method;

[0012] S4: Processing the line mode component of the voltage signal by using an improved complementary set empirical mode decomposition algorithm to obtain an intrinsic mode function of the line mode component of the fault line;

[0013] S5: converting the intrinsic mode function into a feature map by image quantization coding;

[0014] S6: Combine the line mode component of the voltage signal in step S3 and the characteristic signal in step S5 Figure 1 The voltage signal is input into a multi-channel convolutional neural network, and the convolutional neural network is used to mine the line mode component of the voltage signal and the features of the feature graph to obtain the fault distance.

[0015] As a further improvement of the distribution network fault location method based on phase voltage mode conversion and multi-channel data fusion, the transformation matrix T and the reversible matrix T selected by the mode conversion method in step S3 are 1 for:

[0016]

[0017] As a further improvement of the distribution network fault location method based on phase voltage mode transformation and multi-channel data fusion, the improved complementary set empirical mode decomposition algorithm in step S4 specifically includes the following steps:

[0018] S4-1: Standardize the line mode component of the voltage signal to U(t), and convert the K white noise n with a mean of 0 and opposite direction to i (t) and -n i (t) are added to U(t) respectively, that is:

[0019]

[0020] Among them, a i is the amplitude of the noise signal, K is the number of added noise, i = 1, 2, ..., K;

[0021] S4-2: Decomposition using the Empirical Mode Decomposition algorithm and The intrinsic mode function is obtained and Right now:

[0022]

[0023] S4-3: Use the following formula to and The integrated average is c i (t), eliminate white noise as much as possible, that is:

[0024]

[0025] S4-4: Using multi-scale permutation entropy partial mean detection c i (t) is an abnormal signal, if so, it is subtracted from the original signal U(t) to obtain the residual signal r(t);

[0026] S4-5: Perform empirical mode decomposition on r(t), denormalize the result, obtain multiple intrinsic mode functions corresponding to different frequency ranges, and then arrange the intrinsic mode functions in order from high frequency to low frequency.

[0027] As a further improvement of the distribution network fault location method based on phase voltage mode conversion and multi-channel data fusion, step S5 specifically includes:

[0028] The multiple sorted intrinsic mode functions obtained in step S4 are mapped to the pixel value range of [0, 255] by the following formula to generate a feature map with pixel features:

[0029]

[0030] Among them, P(m,n) represents the pixel value of the mth row and nth column of the feature map, x m，n represents the value of the nth data point in the mth intrinsic mode function, x min and x max They represent the maximum and minimum values ​​of all data points in the intrinsic mode function respectively, and round(·) means rounding.

[0031] As a further improvement of the distribution network fault location method based on phase voltage mode transformation and multi-channel data fusion, the method of mining the line mode component of the voltage signal and the features of the feature map using a convolutional neural network in step S6 is:

[0032] S6-1: using a one-dimensional convolutional neural network to mine the line mode component of the voltage signal and extract the global features in the time domain; and using a two-dimensional convolutional neural network to process the feature map and mine the deep features in the spatial domain;

[0033] S6-2: concatenating the global features in the time domain and the deep features in the space domain to form a comprehensive feature vector;

[0034] S6-3: Input the comprehensive feature vector into the third fully connected layer for nonlinear feature learning, and obtain the fault distance by training the network model.

[0035] As a further improvement of the distribution network fault location method based on phase voltage mode transformation and multi-channel data fusion: the one-dimensional convolutional neural network includes a first one-dimensional convolutional layer, a batch normalization layer, a first corrected linear unit activation function layer, a second one-dimensional convolutional layer, a second corrected linear unit activation function layer, a flattening layer and a first fully connected layer connected in sequence.

[0036] As a further improvement of the distribution network fault location method based on phase voltage mode transformation and multi-channel data fusion: the first one-dimensional convolution layer uses 32 5×1 convolution kernels for feature extraction, and the second one-dimensional convolution layer uses 64 3×1 convolution kernels for feature extraction.

[0037] As a further improvement of the distribution network fault location method based on phase voltage mode transformation and multi-channel data fusion: the two-dimensional convolutional neural network adopts a ResNet50 network with a residual structure, and the ResNet50 network includes a two-dimensional convolutional layer, a maximum pooling layer, a residual module group, an average pooling layer and a second fully connected layer connected in sequence.

[0038] As a further improvement of the distribution network fault location method based on phase voltage mode transformation and multi-channel data fusion: the two-dimensional convolution layer uses 64 7×7 convolution kernels for feature extraction, the residual module group includes 16 residual modules connected in sequence, and each residual module consists of two 1×1 convolution layers and one 3×3 convolution layer.

[0039] As a further improvement of the distribution network fault location method based on phase voltage mode transformation and multi-channel data fusion: the number of neurons in the second fully connected layer is 1000.

[0040] Compared with the prior art, the present invention has the following beneficial effects: (1) by using the phase voltage mode transformation method to convert the three-phase voltage signal into the voltage line mode component, the fault information in the distribution network can be extracted more accurately; further using the complementary set empirical mode decomposition method, the extracted intrinsic mode function is converted into a two-dimensional feature map through image quantization coding, which enhances the visualization effect of the signal and the recognition ability of the feature, and improves the accuracy of the distribution network fault distance measurement; (2) the line mode component of the one-dimensional voltage signal and the generated two-dimensional spatial domain feature map are simultaneously input into the multi-channel convolutional neural network for feature mining, which not only makes full use of the one-dimensional voltage signal The timing information of the three-phase voltage signal is combined with the spatial information of the two-dimensional image, thereby realizing a comprehensive analysis and mining of the voltage signal characteristics, further improving the accuracy of fault location. (3) The phase voltage analog conversion and image quantization coding technology optimize the calculation process in the feature extraction process, and has a faster conversion speed than the traditional method. This efficient processing method reduces the calculation time and enables fault location to be carried out more real-time and efficiently. (4) When performing analog conversion on the three-phase voltage signal, the signal noise reduction and smoothing processing and advanced feature conversion methods are used to effectively reduce the influence of noise and interference, thereby improving the robustness and reliability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 is a flow chart of the present invention;

[0042] Figure 2 It is a result schematic diagram of the line mode component of the voltage signal obtained by the mode conversion method;

[0043] Figure 3 This is a schematic diagram of the result of converting the line mode component into a characteristic map;

[0044] Figure 4 This is the structural diagram of the multi-channel convolutional neural network. DETAILED DESCRIPTION

[0045] The technical solution of the present invention is described in detail below with reference to the accompanying drawings:

[0046] like Figure 1 A distribution network fault location method based on phase voltage mode conversion and multi-channel data fusion includes the following steps:

[0047] S1: Start the fault detection criterion by monitoring the zero-sequence voltage on the medium voltage side. Specifically, determine whether the zero-sequence voltage anomaly ΔU0 exceeds the set threshold. If so, it indicates that a ground fault has been detected in the distribution network, and execute step S2. Otherwise, repeat this step.

[0048] Usually, the threshold is set to the phase voltage U nHowever, as the asymmetry of the system changes, the threshold will be adjusted accordingly to ensure the accuracy and reliability of the detection.

[0049] S2: Collect three-phase voltage data of the fault line of the distribution network, and take a period of data as the data to be processed. Preferably, take two periods of data after the fault as the data to be processed.

[0050] S3: Convert the three-phase voltage of the fault line into the line mode component of the voltage signal through the mode conversion method.

[0051] The essence of the phase transformation is to diagonalize the parameter matrix. Different phase transformation choices will affect the effect of fault diagnosis. The present invention adopts a new phase transformation method, which can effectively reflect all types of faults through a single phase component. The selected transformation matrix T and reversible matrix T -1 Such as formula (1) and formula (2).

[0052]

[0053]

[0054] Figure 2 This is a schematic diagram of the line mode component of the voltage signal obtained by the mode conversion method.

[0055] S4: Processing the line mode component of the voltage signal by using an improved complementary set empirical mode decomposition algorithm to obtain an intrinsic mode function of the line mode component of the fault line.

[0056] Specifically, the steps include:

[0057] S4-1: Standardize the line mode component of the voltage signal to U(t), and convert the K white noise n with a mean of 0 and opposite direction to i (t) and -n i (t) are added to U(t) respectively, that is:

[0058]

[0059] Among them, a i is the amplitude of the noise signal, which is selected as 0.2 times the standard deviation of the U(t) signal, K is the number of added noise, i = 1, 2, ..., K;

[0060] S4-2: Decomposition using the Empirical Mode Decomposition algorithm and The intrinsic mode function is obtained and Right now:

[0061]

[0062] S4-3: Use the following formula to and The integrated average is c i (t), eliminate white noise as much as possible, that is:

[0063]

[0064] S4-4: Using multi-scale permutation entropy partial mean detection c i (t) is an abnormal signal, if so, it is subtracted from the original signal U(t) to obtain the residual signal r(t);

[0065] S4-5: Perform empirical mode decomposition on r(t), denormalize the result, obtain multiple intrinsic mode functions corresponding to different frequency ranges, and then arrange the intrinsic mode functions in order from high frequency to low frequency.

[0066] S5: Convert the intrinsic modal function into a feature map by image quantization coding. Specifically, the multiple sorted intrinsic modal functions obtained in step S4 are mapped to a pixel value range of [0, 255] by the following formula (6) to generate a feature map with pixel features:

[0067]

[0068] Among them, P(m,n) represents the pixel value of the mth row and nth column of the feature map, x m，n represents the value of the nth data point in the mth intrinsic mode function, x min and x max They represent the maximum and minimum values ​​of all data points in the intrinsic mode function respectively, and round(·) means rounding.

[0069] In this embodiment, the output of step S4 is 5 intrinsic mode functions. The time-frequency data matrix formed by the sorted 5 intrinsic mode functions can be converted into a pixel matrix, that is, a feature map, through the above method.

[0070] Figure 3 This is a schematic diagram of the result of converting the line mode component into a feature map. This method does not require artificial setting of feature values ​​in advance, and the obtained feature map contains rich point, line, surface and other features, which can reflect the current system operation status.

[0071] Phase voltage mode conversion and image quantization coding technology optimize the calculation process in the feature extraction process, and have a faster conversion speed than traditional methods. This efficient processing method reduces the calculation time and enables fault location to be carried out more real-time and efficiently.

[0072] During the mode conversion process, the signal noise reduction and smoothing processing and advanced feature conversion methods effectively reduce the impact of noise and interference, and improve the robustness and reliability of the system. This means that in a complex power system environment, the fault location results can still maintain a high degree of accuracy and stability.

[0073] S6: Combine the line mode component of the voltage signal in step S3 and the characteristic signal in step S5 Figure 1 The signal is input into a multi-channel convolutional neural network, and the convolutional neural network is used to mine the line mode component of the voltage signal and the features of the feature map to obtain the fault distance.

[0074] Specifically, the steps include:

[0075] S6-1: Use a one-dimensional convolutional neural network (1D-CNN) to mine the line mode component of the voltage signal and extract the global features in the time domain; at the same time, use a two-dimensional convolutional neural network to process the feature map and mine the deep features in the spatial domain.

[0076] Figure 4 It is a structural diagram of a multi-channel convolutional neural network. The one-dimensional convolutional neural network includes a first one-dimensional convolutional layer (Conv1D), a batch normalization layer (BN), a first rectified linear unit (ReLU) activation function layer, a second one-dimensional convolutional layer, a second rectified linear unit activation function layer, a flattening layer (F|atten) and a first fully connected layer connected in sequence. The first one-dimensional convolutional layer uses 32 5×1 convolution kernels for feature extraction, and the second one-dimensional convolutional layer uses 64 3×1 convolution kernels for feature extraction. After the convolution operation, it is processed by the ReLU activation function and converted into the input of 2000 neurons through the flattening layer. The fully connected layer contains 1000 neurons. The introduction of the batch normalization layer is intended to speed up the convergence of the network and prevent gradient disappearance and network overfitting. The application of the ReLU activation function effectively avoids the problem of gradient dissipation, improves the sparsity and nonlinear characteristics of the network, and helps to reduce overfitting. In addition, in order to reduce the classification generalization error, this technology also introduces the Dropout technology, which enhances the robustness of the network by randomly deleting some neurons.

[0077] The two-dimensional convolutional neural network adopts the ResNet50 network with a residual structure, and the ResNet50 network includes a two-dimensional convolutional layer, a maximum pooling layer (3×3), a residual module group, an average pooling layer (7×7) and a second fully connected layer connected in sequence. The two-dimensional convolutional layer uses 64 7×7 convolution kernels for feature extraction, and the residual module group includes 16 residual modules connected in sequence, each of which consists of two 1×1 convolutional layers and one 3×3 convolutional layer. This design helps to increase the dimension of the feature map and alleviate the problems of gradient vanishing and network degradation in deep convolutional neural networks. In order to enable the network to fairly learn the dual-channel feature information of the one-dimensional and two-dimensional processing modules, this patent adjusts the number of neurons in the second fully connected layer of the ResNet50 network from 2048 to 1000.

[0078] S6-2: Concatenate the global features in the time domain and the deep features in the space domain to form a comprehensive feature vector, which integrates the feature information in the time domain and the space domain.

[0079] S6-3: The comprehensive feature vector is input into the third fully connected layer for nonlinear feature learning. The number of neurons in the fully connected layer is set to 128. By training the network model, the distance of the fault point relative to the line head end can be accurately predicted. This method not only utilizes the characteristics of the time domain and the space domain, but also improves the accuracy of fault location and the comprehensive performance of the model through the deep learning ability of the fully connected layer.

[0080] Aiming at the problems existing in the fault distance measurement of distribution network, the present invention proposes a novel fault distance measurement method based on phase voltage mode transformation and multi-channel convolutional neural network. Firstly, the phase voltage mode transformation technology is used to convert the three-phase voltage signal into the voltage line mode component, which is an effective feature for analyzing the transient process of ground fault in distribution network. Subsequently, these voltage line mode components are processed by the improved complementary set empirical mode decomposition method, and a series of intrinsic mode functions are extracted from them. In order to improve the availability and recognition accuracy of these features, the intrinsic mode function is converted into a two-dimensional feature map, and the image quantization coding technology is used to enhance the visualization effect of the features. Finally, the line mode component of the one-dimensional voltage signal and the generated two-dimensional spatial domain feature map are synchronously input into the multi-channel convolutional neural network for processing. In the network, the deep features of the voltage signal and the feature image are respectively mined by the one-dimensional convolution layer and the two-dimensional convolution layer, so as to improve the accuracy of the fault distance measurement of the distribution network. The method has been verified by different topological structure models, proving that it can effectively improve the accuracy and anti-interference ability of fault distance measurement under various operating conditions, and has strong versatility and can be applied to different types of distribution network systems.

Claims

1. A distribution network fault location method based on phase voltage mode conversion and multi-channel data fusion, characterized in that: The steps include: S1: Determine whether the zero-sequence voltage anomaly exceeds a set threshold. If so, it indicates that a ground fault has been detected in the distribution network, and step S2 is executed. Otherwise, this step is repeated. S2: Collect the three-phase voltage data of the fault line of the distribution network, and take a period of cycle data as the data to be processed; S3: converting the three-phase voltage of the fault line into the line mode component of the voltage signal by means of a mode conversion method; S4: Processing the line mode component of the voltage signal by using an improved complementary set empirical mode decomposition algorithm to obtain an intrinsic mode function of the line mode component of the fault line; S5: converting the intrinsic mode function into a feature map by image quantization coding; S6: Input the line mode component of the voltage signal in step S3 and the feature map in step S5 into a multi-channel convolutional neural network, and use the convolutional neural network to mine the features of the line mode component of the voltage signal and the feature map to obtain the fault distance.

2. The distribution network fault location method based on phase voltage mode conversion and multi-channel data fusion according to claim 1 is characterized in that: The transformation matrix T and the reversible matrix T selected by the modular transformation method in step S3 -1 for:

3. The distribution network fault location method based on phase voltage mode conversion and multi-channel data fusion as claimed in claim 1, characterized in that: The improved complementary set empirical mode decomposition algorithm described in step S4 specifically includes the following steps: S4-1: Standardize the line mode component of the voltage signal to U(t), and convert the K white noise n with a mean of 0 and opposite direction to i (t) and -n i (t) are added to U(t) respectively, that is: Among them, a i is the amplitude of the noise signal, K is the number of added noise, i = 1, 2, ..., K; S4-2: Decomposition using the empirical mode decomposition algorithm and The intrinsic mode function is obtained and Right now: S4-3: Use the following formula to and The integrated average is c i (t), eliminate white noise as much as possible, that is: S4-4: Using multi-scale permutation entropy partial mean detection c i (t) is an abnormal signal, if so, it is subtracted from the original signal U(t) to obtain the residual signal r(t); S4-5: Perform empirical mode decomposition on r(t), denormalize the result, obtain multiple intrinsic mode functions corresponding to different frequency ranges, and then arrange the intrinsic mode functions in order from high frequency to low frequency.

4. The distribution network fault location method based on phase voltage mode conversion and multi-channel data fusion according to claim 1, characterized in that: The step S5 specifically includes: The multiple sorted intrinsic mode functions obtained in step S4 are mapped to the pixel value range of [0, 255] by the following formula to generate a feature map with pixel features: Among them, P(m,n) represents the pixel value of the mth row and nth column of the feature map, x m，n represents the value of the nth data point in the mth intrinsic mode function, x min and x max They represent the maximum and minimum values ​​of all data points in the intrinsic mode function respectively, and round(·) means rounding.

5. The distribution network fault location method based on phase voltage mode conversion and multi-channel data fusion as claimed in claim 1, characterized in that: The method of using a convolutional neural network to mine the line mode component of the voltage signal and the features of the feature graph in step S6 is: S6-1: using a one-dimensional convolutional neural network to mine the line mode component of the voltage signal and extract the global features in the time domain; and using a two-dimensional convolutional neural network to process the feature map and mine the deep features in the spatial domain; S6-2: concatenating the global features in the time domain and the deep features in the space domain to form a comprehensive feature vector; S6-3: Input the comprehensive feature vector into the third fully connected layer for nonlinear feature learning, and obtain the fault distance by training the network model.

6. The distribution network fault location method based on phase voltage mode conversion and multi-channel data fusion as claimed in claim 5, characterized in that: The one-dimensional convolutional neural network includes a first one-dimensional convolutional layer, a batch normalization layer, a first rectified linear unit activation function layer, a second one-dimensional convolutional layer, a second rectified linear unit activation function layer, a flattening layer and a first fully connected layer, which are connected in sequence.

7. The distribution network fault location method based on phase voltage mode conversion and multi-channel data fusion as claimed in claim 6, characterized in that: The first one-dimensional convolution layer uses 32 5×1 convolution kernels for feature extraction, and the second one-dimensional convolution layer uses 64 3×1 convolution kernels for feature extraction.

8. The distribution network fault location method based on phase voltage mode conversion and multi-channel data fusion as claimed in claim 5, characterized in that: The two-dimensional convolutional neural network adopts a ResNet50 network with a residual structure, and the ResNet50 network includes a two-dimensional convolutional layer, a maximum pooling layer, a residual module group, an average pooling layer and a second fully connected layer connected in sequence.

9. The distribution network fault location method based on phase voltage mode conversion and multi-channel data fusion as claimed in claim 8, characterized in that: The two-dimensional convolution layer uses 64 7×7 convolution kernels to perform feature extraction, and the residual module group includes 16 residual modules connected in sequence, and each residual module consists of two 1×1 convolution layers and one 3×3 convolution layer.

10. The distribution network fault location method based on phase voltage mode conversion and multi-channel data fusion as claimed in claim 8, characterized in that: The number of neurons in the second fully connected layer is 1000.

Images