Single-phase grounding fault section location method based on one-dimensional convolutional neural network

Abstract

The present invention discloses a method for locating a single-phase grounding fault section based on a one-dimensional convolutional neural network. First, the line is segmented, and a digital fault indicator is used to collect the transient zero-sequence current signal of each section of the line in real time. Secondly, after the collected original zero-sequence current signal is normalized, a preliminary characteristic waveform is constructed, and the first half cycle of the preliminary characteristic waveform is extracted to construct the final characteristic waveform. Finally, the characteristic waveform under the fault condition is input into a trained one-dimensional convolutional neural network, and the fault section is intelligently classified by a classifier to achieve accurate positioning of the single-phase grounding fault. This method can accurately locate the fault section of the single-phase grounding fault, has high detection accuracy and response speed, can improve the efficiency and reliability of fault diagnosis of the power system, and is of great significance to the stability and rapid recovery of the power system.

Description

Technical Field

[0001] The invention relates to a single-phase grounding fault section positioning method based on a one-dimensional convolutional neural network. Background Art

[0002] The distribution network is an important part of the power system. In addition to being responsible for the transmission and distribution of electric energy, it also shoulders the important tasks of ensuring the user's electricity experience, maintaining the safety and stable operation of the power grid, and supporting the national economic development. Therefore, how to achieve safe and stable operation of the distribution network has always been the focus of people's attention. The distribution network has many branch lines, complex wiring, and is often composed of a mixture of overhead lines and cables. Compared with the transmission network, the distribution network has a higher line failure rate, of which single-phase grounding faults account for 80%. If the fault is not isolated in time, the line insulation will be damaged, which will evolve into a more serious fault and even cause the system to collapse. At the same time, when a single-phase grounding fault occurs in the resonant grounding system, due to the compensation effect of the neutral point arc suppression coil, the fault current flowing through the fault point is weak, and the fault characteristics are not easy to be detected, which brings great difficulties to the fault location. Therefore, it is of great significance to study the single-phase grounding fault distance measurement technology of the resonant grounding system.

[0003] At present, the traditional single-phase ground fault location method of resonant grounding system mainly relies on manual experience and threshold setting. It usually judges the fault state by simple feature analysis of current and voltage signals. Although some fault location methods based on machine learning have also emerged, these methods have some shortcomings. The main problems of these methods are: (1) Only analyzing a single signal feature makes it difficult to fully characterize the complex characteristics of nonlinear and non-stationary signals, resulting in insufficient fault feature extraction; (2) The selection of parameter thresholds is highly dependent on manual experience, lacks automatic adjustment capabilities, and cannot adapt to different scenarios and complex working conditions; (3) Machine learning-based methods usually require a large number of samples and high computational costs for modeling, which becomes a bottleneck for such methods.

[0004] Therefore, in view of the shortcomings of the existing technology, it is necessary to design a new method to meet the actual usage needs. Summary of the invention

[0005] Purpose of the invention: In order to fully characterize the difference between the faulty section and the sound section of a single-phase grounding fault in a resonant grounding system and to achieve accurate positioning of the single-phase grounding fault, the present invention proposes a single-phase grounding fault ranging method for a resonant grounding system based on a one-dimensional convolutional neural network.

[0006] To achieve the above object, the technical solution adopted by the present invention is:

[0007] A single-phase grounding fault location method for a resonant grounding system based on a one-dimensional convolutional neural network comprises the following steps:

[0008] (1) Use a digital fault indicator (DFI) to segment the system line and collect the transient zero-sequence current signal measured by the DFI of each line segment in real time;

[0009] (2) After normalizing the collected original zero-sequence current signal, construct a preliminary characteristic waveform by splicing the normalized zero-sequence current at the beginning and end of the line segment. Extract the first half cycle of the preliminary characteristic waveform to construct the final characteristic waveform reflecting the difference between the fault segment and the normal segment;

[0010] (3) The characteristic waveform under the fault condition is input into the trained one-dimensional convolutional neural network, and the fault segment is intelligently classified through the Softmax classifier to achieve accurate positioning of the single-phase grounding fault.

[0011] Specifically, the data in step (1) is obtained as follows:

[0012] The line is divided into multiple logical sections based on the three indicators of branch, length and load. Considering the practical feasibility and installation cost, only enough DFIs are installed for the line sections with high probability of failure and containing important loads. DFI equipment is installed at the starting point and end point of each selected section to monitor the transient zero-sequence current signal of the corresponding section. For long lines or areas with frequent faults, DFI is appropriately added in the middle of the line to shorten the segment length. Each DFI collects the transient zero-sequence current signal in real time through the communication unit and transmits the signal to the control center. When the fault occurs at the end of the line, a virtual detection node (VDN) is introduced. The signal measured by the first DFI along the line is reversely processed and regarded as the measurement data of the VDN.

[0013] The original transient zero-sequence current signal X collected at both ends of the j-th line segment j and Y j , where X j =[x1,x2,...,x n ],Y j =[y1,y2,...,y n ] represents the unnormalized zero-sequence current data.

[0014] Specifically, the line segmentation in step (1) is:

[0015] Based on the topological structure of the distribution line, the main line and the branch line are managed separately. Each branch should be an independently monitored unit so that the fault section can be quickly locked when locating the fault.

[0016] Division by line length Taking into account the protection action range and DFI monitoring capability, long lines should be divided into several sections at fixed intervals to ensure the coverage and accuracy of detection. If the length of a section needs to be adjusted, the length of each section after adjustment is l i satisfy:

[0017] l min ≤l i ≤l max

[0018] Among them, l min and l max are the minimum and maximum allowed lengths of each segment, respectively.

[0019] The load distribution and load importance are comprehensively considered according to the load division. In the line segmentation, the monitoring needs of the high-load area take precedence over the low-load area. The segment length should be shortened in the high-load area, and the segment length can be appropriately extended in the low-load area. The segment length is dynamically adjusted according to the importance and distribution of the load to ensure that the segmentation of the high-load area is more refined.

[0020] The load density after defining the comprehensive weight of the i-th section is:

[0021]

[0022] Among them, ρ i is the load density of the i-th section after comprehensive weighting, P ij is the jth load of the i-th section, σ ij is the weight of the jth load in the i-th section, l i is the length of the i-th segment, and N is the number of segments.

[0023] According to the load density distribution, adjust the segment length to make the load of each segment as uniform as possible to meet the following requirements:

[0024]

[0025] Among them, ρ avg is the average load density of the entire line.

[0026] Specifically, the characteristic waveform extraction described in step (2) is:

[0027] The measured transient zero-sequence current is normalized, and the zero-sequence current signal is normalized:

[0028]

[0029] Among them, X j* and Y j* are the normalized zero-sequence currents at both ends of the jth segment, X j and Y jare the original zero-sequence currents at both ends of the j-th section respectively.

[0030] The normalized zero-sequence current signal X j* and Y j* Splice by segment to construct a preliminary characteristic waveform:

[0031]

[0032] Extract the first half cycle of the preliminary characteristic waveform and construct the final characteristic waveform reflecting the difference between the fault segment and the normal segment;

[0033] Specifically, the design and training of the one-dimensional convolutional neural network described in step (3) is:

[0034] The structure of 1-D CNN consists of five layers: input layer, convolution layer, pooling layer, fully connected layer and output layer.

[0035] In the convolution layer, the specific equation of the convolution operation is:

[0036]

[0037] Among them, l represents the number of network layers, K is the convolution kernel, and b is the bias. represents the output of the lth convolutional layer, is the input of the l-1th layer, M j is the input feature set, and f is the activation function.

[0038] In the pooling layer, the specific equation of the average pooling operation is:

[0039]

[0040] Among them, a and b are weight and bias respectively, and k and t are the dimensions of the pooling matrix respectively.

[0041] In the fully connected layer, the output features of the previous layer are expanded one by one as column vectors and stacked to form a single column feature vector, which is then mapped by the following formula:

[0042] y k =f(ω k ·x+b k )

[0043] Among them, y k is the output of the fully connected layer, ω k and b k are weight and bias respectively.

[0044] The 1-D CNN is trained using historical data under various fault conditions to obtain the optimal convolution kernel parameters, so that the network can better extract local features of the input data.

[0045] Specifically, the precise positioning of the single-phase grounding fault in step (3) is:

[0046] When a single-phase-to-ground fault occurs, the DFI measures the fault signal at a sampling rate of 6.4kHz. The first half cycle of the zero-sequence current is processed to construct characteristic waveforms, which are used as inputs to the trained 1-DCNN to locate the fault section.

[0047] The feature vector generated in step (2) is input into the input layer, passed through the convolution layer, pooling layer and fully connected layer, and the output is input into a two-way Softmax function, which is used as a classifier to generate the classification result of the fault section location:

[0048]

[0049] u 0_ij 、u 1_ij Denotes DFI i and DFI j The measured characteristic waveform belongs to the normal section and the fault section.

[0050] Based on the classification results, normal sections and faulty sections are distinguished, and the specific location of the faulty section is associated with specific distribution line information; the classification results of the faulty section location are stored in the distribution system database, and auxiliary decision support is provided for subsequent possible fault isolation and service restoration.

[0051] An electronic device includes one or more processors for storing one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors execute a method for locating a single-phase grounding fault section in a resonant grounding system based on a 1-D convolutional neural network.

[0052] A storage medium stores a computer program, wherein the computer program is configured to execute a method for locating a single-phase grounding fault section in a resonant grounding system based on a 1-D convolutional neural network when running. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a flow chart of the present invention.

[0054] Figure 2 This is the distribution diagram of DFI in the resonant grounding system.

[0055] Figure 3 This is the structural diagram of a one-dimensional convolutional neural network. DETAILED DESCRIPTION

[0056] The following is combined with Figure 1-3 , the present invention is further described.

[0057] The present invention proposes a single-phase grounding fault section location method based on a one-dimensional convolutional neural network, comprising the following steps:

[0058] To achieve the above object, the technical solution adopted by the present invention is:

[0059] A DC distribution line fault identification method based on improved VMD and SAE includes the following steps:

[0060] (1) Data acquisition. Use digital fault indicators to segment the system line and collect the transient zero-sequence current signals measured by DFI in each section of the line in real time;

[0061] (2) Characteristic waveform extraction. After normalizing the collected original zero-sequence current signal, construct a preliminary characteristic waveform by splicing the normalized zero-sequence current at the beginning and end of the line segment. Extract the first half cycle of the preliminary characteristic waveform to construct the final characteristic waveform that reflects the difference between the fault segment and the normal segment;

[0062] (3) Accurate positioning of single-phase grounding faults. The characteristic waveform under the fault condition is input into the trained one-dimensional convolutional neural network, and the fault segment is intelligently classified through the Softmax classifier to achieve accurate positioning of the single-phase grounding fault.

[0063] Specifically, the data in step (1) is obtained as follows:

[0064] The line is divided into multiple logical sections based on the three indicators of branch, length and load. Considering the practical feasibility and installation cost, sufficient DFI is installed only for the line sections with high probability of failure and containing important loads. DFI equipment is installed at the starting point and end point of each selected section to monitor the transient zero-sequence current signal of the corresponding section. The distribution of DFI in a typical resonant grounding system is shown in the figure below. Figure 2 As shown. For long lines or fault-prone areas, DFIs are appropriately added in the middle of the line to shorten the segment length. Each DFI collects transient zero-sequence current signals in real time through a communication unit and transmits the signals to the control center. When a fault occurs at the end of the line, a virtual detection node (VDN) is introduced. The signal measured by the first DFI along the line is reversely processed and regarded as the measurement data of the VDN.

[0065] The original transient zero-sequence current signal X collected at both ends of the j-th line segment j and Y j , where X j =[x1,x2,...,x n ],Y j =[y1,y2,...,y n ] represents the unnormalized zero-sequence current data.

[0066] Specifically, the line segmentation in step (1) is:

[0067] Based on the topological structure of the distribution line, the main line and the branch line are managed separately. Each branch should be an independently monitored unit so that the fault section can be quickly locked when locating the fault.

[0068] Division by line length Taking into account the protection action range and DFI monitoring capability, long lines should be divided into several sections at fixed intervals to ensure the coverage and accuracy of detection. If the length of a section needs to be adjusted, the length of each section after adjustment is l i satisfy:

[0069] l min ≤l i ≤l max

[0070] Among them, l min and l max are the minimum and maximum allowed lengths of each segment, respectively.

[0071] The load distribution and load importance are comprehensively considered according to the load division. In the line segmentation, the monitoring needs of the high-load area take precedence over the low-load area. The segment length should be shortened in the high-load area, and the segment length can be appropriately extended in the low-load area. The segment length is dynamically adjusted according to the importance and distribution of the load to ensure that the segmentation of the high-load area is more refined.

[0072] The load density after defining the comprehensive weight of the i-th section is:

[0073]

[0074] Among them, ρ i is the load density of the i-th section after comprehensive weighting, P ij is the jth load of the i-th section, σ ij is the weight of the jth load in the i-th section, l i is the length of the i-th segment, and N is the number of segments.

[0075] According to the load density distribution, adjust the segment length to make the load of each segment as uniform as possible to meet the following requirements:

[0076]

[0077] Among them, ρ avg is the average load density of the entire line.

[0078] Specifically, the characteristic waveform extraction described in step (2) is:

[0079] The measured transient zero-sequence current is normalized, and the zero-sequence current signal is normalized:

[0080]

[0081] Among them, X j* and Y j* are the normalized zero-sequence currents at both ends of the jth segment, X j and Y j are the original zero-sequence currents at both ends of the j-th section respectively.

[0082] The normalized zero-sequence current signal X j* and Y j* Splice by segment to construct a preliminary characteristic waveform:

[0083]

[0084] Extract the first half cycle of the preliminary characteristic waveform and construct the final characteristic waveform reflecting the difference between the fault segment and the normal segment;

[0085] Specifically, the design and training of the one-dimensional convolutional neural network described in step (3) is:

[0086] The structure of 1-DCNN includes five layers: input layer, convolution layer, pooling layer, fully connected layer and output layer. Figure 3 shown.

[0087] In the convolution layer, the specific equation of the convolution operation is:

[0088]

[0089] Among them, l represents the number of network layers, K is the convolution kernel, and b is the bias. represents the output of the lth convolutional layer, is the input of the l-1th layer, M j is the input feature set, and f is the activation function.

[0090] In the pooling layer, the specific equation of the average pooling operation is:

[0091]

[0092] Among them, a and b are weight and bias respectively, and k and t are the dimensions of the pooling matrix respectively.

[0093] In the fully connected layer, the output features of the previous layer are expanded one by one as column vectors and stacked to form a single column feature vector, which is then mapped by the following formula:

[0094] y k =f(ω k ·x+b k )

[0095] Among them, yk is the output of the fully connected layer, ω k and b k are weight and bias respectively.

[0096] The 1-D CNN is trained using historical data under various fault conditions to obtain the optimal convolution kernel parameters, so that the network can better extract local features of the input data.

[0097] Specifically, the precise positioning of the single-phase grounding fault in step (3) is:

[0098] When a single-phase-to-ground fault occurs, the DFI measures the fault signal at a sampling rate of 6.4kHz. The first half cycle of the zero-sequence current is processed to construct characteristic waveforms, which are used as inputs to the trained 1-DCNN to locate the fault section.

[0099] The feature vector generated in step (2) is input into the input layer, passed through the convolution layer, pooling layer and fully connected layer, and the output is input into a two-way Softmax function, which is used as a classifier to generate the classification result of the fault section location:

[0100]

[0101] u 0_ij 、u 1_ij Denotes DFI i and DFI j The measured characteristic waveform belongs to the normal section and the fault section.

[0102] Based on the classification results, normal sections and faulty sections are distinguished, and the specific location of the faulty section is associated with specific distribution line information; the classification results of the faulty section location are stored in the distribution system database, and auxiliary decision support is provided for subsequent possible fault isolation and service restoration.

[0103] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A single-phase grounding fault section location method based on one-dimensional convolutional neural network, characterized by: The following steps are involved: (1) Use digital fault indicators to segment the system line and collect the transient zero-sequence current signals measured by DFI in each section of the line in real time; (2) After normalizing the collected original zero-sequence current signal, a preliminary characteristic waveform is constructed by splicing the normalized zero-sequence currents at the beginning and end of the line segment; the first half cycle of the preliminary characteristic waveform is extracted to construct a final characteristic waveform reflecting the difference between the fault segment and the normal segment; (3) The characteristic waveform under the fault condition is input into the trained one-dimensional convolutional neural network, and the fault segment is intelligently classified through the Softmax classifier to achieve accurate positioning of the single-phase grounding fault.

2. According to the method for locating a single-phase grounding fault section based on a one-dimensional convolutional neural network as described in claim 1, it is characterized in that: The step (1) specifically comprises: The line is divided into multiple logical sections based on the three indicators of branch, length and load. Considering the practical feasibility and installation cost, sufficient DFIs are installed only for line sections with high probability of failure and important loads. DFI equipment is installed at the starting point and end point of each selected section to monitor the transient zero-sequence current signal of the corresponding section. For long lines or areas with frequent faults, DFI is appropriately added in the middle of the line to shorten the segment length. Each DFI collects transient zero-sequence current signals in real time through a communication unit and transmits the signals to the control center. When a fault occurs at the end of the line, a virtual detection node is introduced. The signal measured by the first DFI along the line is reversely processed and regarded as the measurement data of the VDN. The original transient zero-sequence current signal X collected at both ends of the j-th line segment j and Y j , where X j =[x1,x2,...,x n ],Y j =[y1,y2,...,y n ] represents the unnormalized zero-sequence current data.

3. The line segmentation method according to claim 2, characterized in that: The line segmentation principle is: Based on the topological structure of the distribution line, the main line and the branch line are managed separately; each branch should be an independently monitored unit so that the fault section can be quickly locked when locating the fault; Division by line length Taking into account the protection action range and DFI monitoring capability, long lines are divided into several sections at fixed intervals to ensure the coverage and accuracy of detection; if the length of a section needs to be adjusted, the length of each section after adjustment is l i satisfy: l min ≤l i ≤l max Among them, l min and l max are the minimum and maximum allowed lengths of each segment, respectively.

4. The line segmentation method according to claim 3, characterized in that: The load distribution and load importance are comprehensively considered according to the load division; in the line segmentation, the monitoring needs of the high-load area take precedence over the low-load area. The segment length should be shortened in the high-load area, and the segment length can be appropriately extended in the low-load area; the segment length is dynamically adjusted according to the importance and distribution of the load to ensure that the segmentation of the high-load area is more refined; The load density after defining the comprehensive weight of the i-th section is: Among them, ρ i is the load density of the i-th section after comprehensive weighting, P ij is the jth load of the i-th section, σ ij is the weight of the jth load in the i-th section, l i is the length of the i-th segment, and N is the number of segments; According to the load density distribution, adjust the segment length to make the load of each segment as uniform as possible to meet the following requirements: Among them, ρ avg is the average load density of the entire line.

5. According to the method for locating a single-phase grounding fault section based on a one-dimensional convolutional neural network as described in claim 1, it is characterized in that: The step (2) is specifically: The measured transient zero-sequence current is normalized, and the zero-sequence current signal is normalized: Among them, X j* and Y j* are the normalized zero-sequence currents at both ends of the jth segment, X j and Y j are the original zero-sequence currents at both ends of the j-th section respectively.

6. A method for locating a single-phase grounding fault section based on a one-dimensional convolutional neural network according to claim 5, characterized in that: The normalized zero-sequence current signal X j* and Y j* Splice by segment to construct a preliminary characteristic waveform: The first half cycle of the preliminary characteristic waveform is extracted to construct the final characteristic waveform reflecting the difference between the fault segment and the normal segment.

7. According to a method for locating a single-phase grounding fault section based on a one-dimensional convolutional neural network as described in claim 1, it is characterized in that: The step (3) specifically comprises: When a single-phase-to-ground fault occurs, the DFI measures the fault signal at a sampling rate of 6.4kHz. The first half cycle of the zero-sequence current is processed to construct characteristic waveforms, which are used as inputs to the trained 1-D CNN to locate the fault section. The feature vector generated in step (2) is input into the input layer, passed through the convolution layer, pooling layer and fully connected layer, and the output is input into a two-way Softmax function, which is used as a classifier to generate the classification result of the fault section location: u 0_ij 、u 1_ij Denotes DFI i and DFI j The membership degree of the measured characteristic waveform to the normal section and the fault section; Based on the classification results, normal sections and faulty sections are distinguished, and the specific location of the faulty section is associated with specific distribution line information; the classification results of the faulty section location are stored in the distribution system database, and auxiliary decision support is provided for subsequent possible fault isolation and service restoration.

8. An electronic device, characterized in that: It includes one or more processors for storing one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors execute the single-phase grounding fault section locating method based on a one-dimensional convolutional neural network as described in any one of claims 1 to 7.

9. A storage medium storing a computer program, wherein: in, The computer program is configured to execute the single-phase grounding fault section location method based on a one-dimensional convolutional neural network as described in any one of claims 1 to 5 when running.

Images