Fault detection method for high-resistance grounding fault of power distribution network

Abstract

The invention specifically relates to a fault detection method for a high-resistance grounding fault of a power distribution network, and belongs to the technical field of power distribution network fault identification. According to the method, the differentiated change characteristics of the long-time scale zero-sequence voltage waveform under fault and disturbance events are utilized, and the characteristic distinguishability is enhanced through image processing; a fast regional convolutional neural network is introduced to establish a power distribution network high-resistance grounding fault detection model, and end-to-end fault classification and positioning are realized. Finally, the initial fault moment is corrected in combination with a change point detection algorithm, and the positioning precision is improved.

Description

Technical Field

[0001] This invention specifically relates to a fault detection method for high-resistance grounding faults in distribution networks, belonging to the field of distribution network fault identification technology. Background Technology

[0002] Over 90% of power outages in power systems are caused by distribution line faults. Single-phase grounding faults are the most common, accounting for over 80% of all faults. High-resistance grounding faults are a special type of single-phase grounding fault, generally caused by events such as lightning strikes or strong winds causing overhead lines to break and fall to the ground, coming into contact with a high-resistance grounding medium. Common high-resistance grounding media include cement, sand, branches, and grass, with resistances ranging from hundreds to tens of thousands of ohms. High-resistance grounding faults are usually accompanied by electric arcing; if not promptly extinguished, they can pose significant safety hazards, such as fires, equipment damage, and even personal injury. Therefore, accurately identifying high-resistance grounding faults is crucial for maintaining power grid safety and ensuring power supply reliability.

[0003] Existing methods mainly rely on threshold settings or manual feature extraction, which suffers from poor adaptability and susceptibility to disturbances. Although deep learning technology has been applied in fault diagnosis, most methods are limited to short-term data and struggle to distinguish between permanent faults and transient disturbances. Summary of the Invention

[0004] The technical problem to be solved by this invention is: how to improve the detection accuracy of high-resistance grounding faults in power distribution networks.

[0005] To solve the above-mentioned technical problems, the technical solution proposed by this invention is: a fault detection method for high-resistance grounding faults in distribution networks, comprising the following steps:

[0006] Step 1: Establish an initial detection model for high-resistivity grounding faults in the distribution network based on a fast regional convolutional neural network;

[0007] Step 2: Collect n segments of historical high-resistance grounding fault zero-sequence voltage signals and m segments of historical disturbance event zero-sequence voltage signals during the distribution network's historical process.

[0008] Step 3: Preprocess the n segments of historical high-resistance ground fault zero-sequence voltage signals and the m segments of historical disturbance event zero-sequence voltage signals, and convert the preprocessed zero-sequence voltage signals into grayscale images; mark the bounding boxes of the converted n historical high-resistance ground fault zero-sequence voltage waveform grayscale images and m historical disturbance event zero-sequence voltage waveform grayscale images.

[0009] Step 4: Input n historical high-resistivity grounding fault grayscale images with bounding boxes and m historical disturbance event grayscale images into the initial detection model of high-resistivity grounding fault in the distribution network for model training to obtain the trained high-resistivity grounding fault detection model of the distribution network; the trained high-resistivity grounding fault detection model of the distribution network takes the grayscale image converted from the zero-sequence voltage signal of the distribution network as input and the grayscale image with bounding boxes containing the location of the fault at the time of occurrence and the fault judgment type as output;

[0010] Step 5: Real-time zero-sequence voltage signal is acquired when the distribution network needs to be detected for high-resistance grounding faults. The real-time zero-sequence voltage signal is preprocessed and converted into a real-time zero-sequence voltage waveform grayscale image. The real-time zero-sequence voltage waveform grayscale image is input into the trained distribution network high-resistance grounding fault detection model to obtain a fault judgment grayscale image with a boundary box containing the fault location and fault judgment type to complete the detection of high-resistance grounding faults in the distribution network.

[0011] Furthermore, the fault detection method also includes the following steps:

[0012] Step 6: Extract image data from three cycles before and after the location of the fault occurrence from the grayscale image obtained in Step 5 as a signal sequence, and use the signal sequence to correct the location of the fault occurrence in the grayscale image. The specific steps are as follows:

[0013] Step 6.1: Initialize the segmentation. Set an initial segmentation point in the signal sequence to divide the complete signal into two sub-segments, the first and the last.

[0014] Step 6.2: Calculate the root mean square value of all data points in the first and second sub-segments respectively, as the baseline statistical characteristics of each segment;

[0015] Step 6.3: For each sub-segment, calculate the deviation between the observed value at that point and the root mean square value of the sub-segment, and further calculate the root mean square deviation of all points in each sub-segment;

[0016] Step 6.4: Add the root mean square deviations of the two sub-segments to obtain the total residual. Iteratively adjust the position of the split point, repeatedly executing steps 6.2 to 6.4 until the total residual reaches the global minimum; the corresponding split point at this time is the location of the fault occurrence after correction.

[0017] Furthermore, the fast region convolutional neural network uses ResNet50 residual network and feature pyramid network as backbone network, and also includes region proposal network, interest region pooling module and fully connected classification and regression layer;

[0018] The ResNet50 residual network and feature pyramid network take grayscale images as input and feature sequences extracted from grayscale images as output.

[0019] The region proposal network uses a sliding window mechanism to slide sample the feature sequences extracted by the ResNet50 residual network and the feature pyramid network, and uses 3×3 convolution kernels to extract local features from the sampling results, generating anchor boxes of different sizes with labels on the grayscale image.

[0020] The region of interest pooling module uniformly fixes anchor frames of different sizes containing labels on the grayscale image into feature blocks of a fixed size;

[0021] The fully connected classification regression layer receives the grayscale image output by the region of interest pooling module and performs classification on the result.

[0022] Furthermore, the preprocessing in step 3 specifically employs a minimum-maximum normalization formula to scale the amplitude of the zero-sequence voltage signal to the range of [0,1]. The specific formula is as follows:

[0023] x norm =(x max -x min ) / (x−x min )

[0024] Where x and x norm These are the amplitude of the signal to be processed and the amplitude of the processed signal in the zero-sequence voltage signal sequence, respectively; x max and x min These are the maximum and minimum amplitudes in the zero-sequence voltage signal sequence, respectively.

[0025] Furthermore, the bounding box annotation process in step 3 is as follows:

[0026] The bounding box annotation of the grayscale image of the zero-sequence voltage waveform of the historical high-resistance grounding fault is defined as follows: the bounding box starts from the moment the fault occurs and covers the range from the rise of the waveform amplitude to the steady state.

[0027] The bounding box of the grayscale image of the zero-sequence voltage waveform of the historical disturbance event is marked, and the bounding box covers the oscillation decay process.

[0028] The beneficial effects of this invention are as follows: This invention utilizes the differentiated change characteristics of long-time scale zero-sequence voltage waveforms under fault and disturbance events, and enhances the distinguishability of features through image processing; This invention also introduces a fast regional convolutional neural network to establish a high-resistance grounding fault detection model for distribution networks, realizing end-to-end fault classification and location; Finally, this invention combines a change point detection algorithm to correct the initial fault time, improving the location accuracy. Attached Figure Description

[0029] Figure 1 This is a flowchart of the fault identification process in an embodiment of the present invention.

[0030] Figure 2 This is a network structure diagram of a fast region convolutional neural network in an embodiment of the present invention. Detailed Implementation

[0031] The following description, in conjunction with the accompanying drawings and specific embodiments, further illustrates the fault detection method for ground faults in a distribution network according to the present invention.

[0032] Example

[0033] The fault detection method in this embodiment, such as Figure 1 As shown, it includes the following steps:

[0034] Step 1: Establish an initial detection model for high-resistivity grounding faults in the distribution network based on a Faster region-based convolutional neural network (Faster R-CNN);

[0035] like Figure 2 As shown, the Fast Region Convolutional Neural Network uses ResNet50 residual network and feature pyramid network as backbone networks, and also includes region proposal network, interest region pooling module and fully connected classification and regression layer;

[0036] The ResNet50 residual network and feature pyramid network take grayscale images as input and feature sequences extracted from grayscale images as output.

[0037] The region proposal network uses a sliding window mechanism to slide sample the feature sequences extracted by the ResNet50 residual network and the feature pyramid network, and uses 3×3 convolution kernels to extract local features from the sampling results, generating anchor boxes of different sizes with labels on the grayscale image.

[0038] The region of interest pooling module unifies and fixes anchor boxes of different sizes containing labels on grayscale images into feature patches of a fixed size;

[0039] The fully connected classification and regression layer receives the grayscale image output by the region of interest pooling module and performs classification on the result.

[0040] Step 2: Collect n segments of historical high-resistance grounding fault zero-sequence voltage signals and m segments of historical disturbance event zero-sequence voltage signals during the distribution network's historical process.

[0041] Step 3: Preprocess the n historical high-resistance ground fault zero-sequence voltage signals and m historical disturbance event zero-sequence voltage signals, and convert the preprocessed zero-sequence voltage signals into grayscale images; mark the bounding boxes of the converted n historical high-resistance ground fault zero-sequence voltage waveform grayscale images and m historical disturbance event zero-sequence voltage waveform grayscale images.

[0042] The preprocessing process specifically uses a minimum-maximum normalization formula to scale the amplitude of the zero-sequence voltage signal to the range of [0,1]. The specific formula is as follows:

[0043] x norm =(x max -x min ) / (x−x min )

[0044] Where x and x norm These are the amplitude of the signal to be processed and the amplitude of the processed signal in the zero-sequence voltage signal sequence, respectively; x max and x min These are the maximum and minimum amplitudes in the zero-sequence voltage signal sequence, respectively.

[0045] The bounding box annotation process is as follows:

[0046] The bounding box annotation of the grayscale image of the zero-sequence voltage waveform of historical high-resistivity ground faults. The bounding box starts from the moment the fault occurs and covers the range from the rise of the waveform amplitude to the steady state.

[0047] The bounding box annotation of the grayscale image of the zero-sequence voltage waveform of historical disturbance events covers the oscillation decay process.

[0048] Step 4: Input n historical high-resistivity grounding fault grayscale images with bounding box annotations and m historical disturbance event grayscale images into the initial high-resistivity grounding fault detection model of the distribution network for model training to obtain the trained distribution network high-resistivity grounding fault detection model; the trained distribution network high-resistivity grounding fault detection model takes the grayscale image converted from the zero-sequence voltage signal of the distribution network as input and the grayscale image with bounding boxes containing the location of the fault at the time of occurrence and the fault judgment type as output;

[0049] Step 5: Real-time zero-sequence voltage signal is acquired when high-resistance grounding fault detection is required in the distribution network. The real-time zero-sequence voltage signal is preprocessed and converted into a real-time zero-sequence voltage waveform grayscale image. The real-time zero-sequence voltage waveform grayscale image is input into the trained distribution network high-resistance grounding fault detection model to obtain a fault judgment grayscale image with the fault occurrence point and fault judgment type bounding box to complete the detection of high-resistance grounding fault in the distribution network.

[0050] Step 6: Extract image data from three cycles before and after the location of the fault occurrence from the grayscale image obtained in Step 5 as a signal sequence. Use this signal sequence to correct the location of the fault occurrence in the grayscale image. The specific steps are as follows:

[0051] Step 6.1: Initialize the segmentation. Set an initial segmentation point in the signal sequence to divide the complete signal into two sub-segments, the first and the last.

[0052] Step 6.2: Calculate the root mean square value of all data points in the first and second sub-segments respectively, as the baseline statistical characteristics of each segment;

[0053] Step 6.3: For each sub-segment, calculate the deviation between the observed value at that point and the root mean square value of the sub-segment, and further calculate the root mean square deviation of all points in each sub-segment;

[0054] Step 6.4: Add the root mean square deviations of the two sub-segments to obtain the total residual. Iteratively adjust the position of the split point, repeatedly executing steps 6.2 to 6.4 until the total residual reaches the global minimum; the corresponding split point at this time is the location of the fault occurrence after correction.

[0055] The above description is only a preferred embodiment of the present invention. All equivalent changes and modifications made within the scope of the claims of the present invention should be included in the scope of the present invention.

Claims

1. A method for detecting a fault of a high impedance ground fault of a power distribution network, characterized in that: The method comprises the following steps: Step 1: establishing an initial detection model for high-resistance grounding faults in a power distribution network according to a fast regional convolutional neural network; Step 2: collecting n historical high-resistance grounding fault zero sequence voltage signals and m historical disturbance event zero sequence voltage signals in a historical process of the power distribution network; Step 3: preprocessing the n historical high-resistance grounding fault zero sequence voltage signals and the m historical disturbance event zero sequence voltage signals, and converting the preprocessed zero sequence voltage signals into gray scale images; performing boundary box labeling on the converted n historical high-resistance grounding fault zero sequence voltage waveform gray scale images and m historical disturbance event zero sequence voltage waveform gray scale images; Step 4: inputting the n historical high-resistance grounding fault gray scale images and m historical disturbance event gray scale images with boundary box labeling into the initial detection model for high-resistance grounding faults in the power distribution network to perform model training, and obtaining a trained high-resistance grounding fault detection model for the power distribution network; the trained high-resistance grounding fault detection model for the power distribution network takes gray scale images converted from power distribution network zero sequence voltage signals as input, and takes gray scale images with boundary boxes of fault time occurrence points and fault judgment types as output; Step 5: collecting real-time zero sequence voltage signals of the power distribution network in need of high-resistance grounding fault detection, preprocessing the real-time zero sequence voltage signals, converting the real-time zero sequence voltage signals into real-time zero sequence voltage waveform gray scale images, and inputting the real-time zero sequence voltage waveform gray scale images into the trained high-resistance grounding fault detection model for the power distribution network to obtain fault judgment gray scale images with boundary boxes of fault time occurrence points and fault judgment types, so as to complete detection of the high-resistance grounding fault in the power distribution network. The fault detection method further comprises the following steps:

2. The fault detection method of claim 1, wherein: Step 6: cutting image data of three cycles before and after the fault time occurrence point from the fault judgment gray scale image of step 5 as a signal sequence, and correcting the fault time occurrence point in the fault judgment gray scale image through the signal sequence, and the specific content is as follows: Step 6.1: initialization segmentation, setting an initial segmentation point in the signal sequence to divide the complete signal into two sub-sections of front and back; Step 6.2: calculating the root mean square values of all data points in the front and back two sub-sections respectively as the reference statistical characteristics of each section; Step 6.3: calculating the deviation between the observation value of each point and the root mean square value of the sub-section for each sub-section, and further calculating the root mean square deviation of all points in each sub-section; Step 6.4: adding the root mean square deviations of the two sub-sections to obtain the total residual error. The position of the segmentation point is adjusted iteratively, and steps 6.2 to 6.4 are repeatedly executed until the total residual error reaches a global minimum value; at this time, the corresponding segmentation point is the corrected fault time occurrence point. The fast regional convolutional neural network adopts a ResNet50 residual network and a feature pyramid network as a backbone network, and further comprises a regional proposal network, an interest region pooling module and a fully connected classification regression layer; 3. The fault detection method of claim 1, wherein: The ResNet50 residual network and the feature pyramid network take gray scale images as input and take feature sequences extracted from the gray scale images as output; ​ The region proposal network adopts a sliding window mechanism to slide sample the feature sequence extracted by the ResNet50 residual network and the feature pyramid network, and uses a 3*3 convolution kernel to extract local features from the sampling results to generate anchor boxes of different sizes containing labels on the gray image; The interest region pooling module fixes the anchor boxes of different sizes containing labels on the gray image to fixed-size feature map blocks; The fully connected classification regression layer accepts the gray image output by the interest region pooling module and classifies the results.

4. The fault detection method of claim 1, wherein: The preprocessing process in step 3 specifically uses the minimum-maximum normalization formula to scale the amplitude of the zero sequence voltage signal to the range of [0, 1], and the specific formula is as follows: x norm =(x max −x min ) / (x−x min ) Wherein, x and x norm are the amplitude of the signal to be processed and the amplitude of the processed signal in the sequence of zero sequence voltage signals, respectively; x max and x min are the maximum and minimum amplitudes in the sequence of zero sequence voltage signals, respectively.

5. The fault detection method of claim 1, wherein: The boundary box labeling process in step 3 is as follows: The boundary box labeling of the historical high-resistance grounding fault zero sequence voltage waveform gray image, the boundary box starts from the fault occurrence time and covers the interval where the waveform amplitude rises to the steady state; The boundary box labeling of the historical disturbance event zero sequence voltage waveform gray image, the boundary box covers the oscillation decay process.

Images