New distribution network fault cell segment positioning method based on time sequence data driving under fault self-healing configuration
By using the Transformer model driven by multi-point traveling wave timing data in the distribution network, the problem of faulty cell segment positioning under the influence of complex topological structures and environmental interference is solved, and the rapid and accurate positioning of distribution network faults is achieved, and the faults are improved. Self-healing ability.
Patent Information
- Application Number
- CN202410948519.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-16
- Publication Date
- 2024-11-08
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
In the distribution network, complex topological structures and environmental interferences lead to difficulty in accurately extracting multi-point traveling wave timing information, and existing fault positioning methods are difficult to achieve accurate fault cell positioning in high-voltage power grid environments.
Using a fault-oriented self-healing configuration method, multi-point traveling wave timing data driven, combined with Transformer neural network model, sample sets are constructed and model evaluation and verification are carried out to ensure high data utilization and accurate results.
It realizes rapid and accurate positioning of faulty cell sections of the distribution network, improves the accuracy and stability of fault positioning, and is suitable for distribution networks with complex topological structures and multiple environmental interferences.
Smart Images

Figure CN118914743A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a new method for locating a distribution network fault section based on time series data under fault self-healing configuration, which realizes rapid positioning of the distribution network fault section by combining multi-point traveling wave time series data in the distribution network with a Transformer model for fault location. Background Art
[0002] The distribution network is an important supporting part of the modern power system. With the development of my country's economy in recent years, the power system needs better operation stability. How to quickly restore the operation of the distribution network after a fault occurs plays an important auxiliary role in the stable operation of the distribution network. The basis of self-healing of distribution network faults is the ability to achieve rapid fault location. There is no doubt that the research on related issues and technologies has profound contemporary significance. The current mainstream distribution network fault location method is the traveling wave method. When a fault occurs in the distribution network, a traveling wave propagating along the line will be generated at the fault point. It has a certain wave speed. The distance to the fault point can be inferred based on the time it takes for the traveling wave signal to reach the traveling wave acquisition device.
[0003] When a fault occurs in the power grid, the traveling wave information of the fault point is unique. It is an important idea to solve this problem by configuring the power grid with multi-point traveling waves and analyzing the multi-point traveling wave time series data of each fault point to locate the fault. The first guarantee for the implementation of this method is a reliable traveling wave time series information acquisition unit. At present, the relevant hardware technology for multi-point traveling wave configuration has been developed and mature. The relevant equipment can accurately obtain the head time of the fault traveling wave and can collect time series data in time through efficient communication means (Li Xinhai, Xiao Xing, Chen Qingjiang, et al. Fault distance measurement method, device and storage medium based on traveling waves [P]. Guangdong Province: CN202410379871.2, 2024-06-21.). At present, many studies have used multi-point traveling wave time series data as the basis for fault location, configured multi-point traveling wave acquisition units for the power grid, and successfully realized fault location using the collected time series information (Yang Jianming, Zhang Qingwu, Wang Yangzheng, et al. Optimization of voltage mutation protection for ultra-high voltage DC transmission lines [J]. Electric Power Engineering Technology, 2022, 41(03): 102-109.). However, most of these studies are based on high-voltage transmission line environments. Their limitations are that the high-voltage power grid structure is simple and intuitive, the environmental interference is small, and the traveling wave information extraction is convenient. In contrast, the distribution network has a complex topological structure and its operating conditions are greatly affected by the environment. There are a series of technical difficulties in accurately extracting multi-point traveling wave time series information when a fault occurs. For the location of distribution network fault sub-sections based on multi-point traveling wave time series information, advanced data-driven technology combined with deep learning neural network models can be used to achieve high-precision distribution network fault sub-section location. Summary of the invention
[0004] In order to achieve rapid self-healing of distribution network faults, the present invention proposes a new method for locating distribution network fault sections based on time series data under fault self-healing configuration. This method mainly aims at the problem of determining the specific location of the fault point in distribution network fault self-healing, and proposes a method for locating distribution network faults by combining distribution network multi-point traveling wave time series data and a Transformer model for fault location.
[0005] The technical solution adopted by the present invention is: a new method for locating a distribution network fault section based on time series data under fault self-healing configuration, the method includes a multi-point traveling wave measurement unit configuration structure for fault self-healing, sample construction driven by time series data, Transformer model construction for fault location, model evaluation indicators, and a model verification method based on critical areas. The specific steps are as follows:
[0006] Step 1: Configure the multi-point traveling wave acquisition unit of the distribution network to ensure that the fault in any section of the distribution network meets the observability of the multi-point traveling wave positioning principle.
[0007] Step 2: Based on the traveling wave time series data collected by the traveling wave acquisition unit, a time series data-driven sample construction is performed.
[0008] Step 3: Build a Transformer neural network model for fault location.
[0009] Step 4: Propose model evaluation indicators to evaluate the performance of the constructed model.
[0010] Step 5: Propose a model verification method based on critical regions to test whether the model performance achieves the desired effect.
[0011] The beneficial effects of the present invention are: (1) This method uses multi-point traveling wave time series data as the basis for fault location, which is more accurate than other methods; (2) The data-driven sample construction method used in this method can ensure high data utilization and more accurate results; (3) This method uses the Transformer neural network model to train the data set, which can ensure the high accuracy of the results. BRIEF DESCRIPTION OF THE DRAWINGS
[0012] Figure 1 The process of locating the faulty small section of the distribution network driven by time series data under fault self-healing configuration.
[0013] Figure 2 Internal structure of the Transformer model.
[0014] Figure 3 10KV distribution network topology diagram.
[0015] Figure 4 Comparison of accuracy with different numbers of attention heads.
[0016] Figure 5 Schematic diagram of the Transformer network confusion matrix.
[0017] Figure 6 Schematic diagram of the Transformer network result distribution.
[0018] Figure 7 Schematic diagram of the Transformer network confusion matrix (method one).
[0019] Figure 8 Schematic diagram of the Transformer network result distribution (method one).
[0020] Fig. 9 Schematic diagram of the Transformer network confusion matrix (method 2).
[0021] Fig.10 Schematic diagram of the Transformer network result distribution (method 2).
[0022] Fig.11 Schematic diagram of the Transformer network confusion matrix (method three).
[0023] Fig.12 Schematic diagram of the Transformer network result distribution (method three). DETAILED DESCRIPTION
[0024] The present invention provides a new method for locating a distribution network fault section based on time series data driven under fault self-healing configuration, which specifically includes the following steps: Step 1: Configure multi-point traveling wave measurement units for fault self-healing. Accurate positioning of the fault section is the basis of fault self-healing, and its hardware configuration should be unified. Considering the complexity of the distribution network topology, its configuration requires that any section of the distribution network meets the observability of the multi-point traveling wave positioning principle. With the development of remote communication technology and high-precision timing systems based on GPS, a positioning system based on wide-area fault traveling wave information is established. The positioning units in the network can be networked and operated, and the main station performs positioning calculations. The information of each measuring point can be integrated to ensure that any fault line in the entire distribution network can be located. Set up acquisition units at the main nodes in the distribution network to collect the traveling wave timing data of the fault point.
[0025] Step 2: Construct samples driven by time series data. The comprehensiveness and generalization of samples are the basis of data-driven methods. For fault location in distribution network sections, the sample construction steps are as follows: a. Configure the collection devices numbered 1 to n for fault self-healing. When the system satisfies U0>U set When the judgment is started, the arrival time of the traveling wave head of each configuration unit is collected from t1 to t n Construct a time series sample. Where n is the number of configured acquisition devices, U0 is the zero sequence voltage, U set is the zero-sequence voltage threshold for fault judgment. The elements in the constructed time series data sample are time series information matrices, denoted by A x , where x is the number of fault points, A x As shown in formula (1): A x =[t1,t2,t3…t n ](1) b. Divide the distribution network line into a small fault section every s kilometers. Considering the positioning accuracy of the fault section, the range of the small section is s≤1km. Set a fault point every t kilometers in each section, simulate different fault types, transition resistances, and fault initial angles, and construct a time series sample set.
[0026] Where s is the length of the fault cell segment and t is the distance between the fault points.
[0027] c. Considering the time loss, anomalies and changes in network topology during normal operation, the model sample data is replaced with 0 values, that is, some data points in the model sample data are replaced with 0 to simulate the above abnormal situations, modify interference factors such as error time, expand the sample set, and improve the generalization ability of the model.
[0028] d. Divide the sample set into N mutually exclusive sub-sample sets of the same size. Randomly select p sample sets for training and q sample sets for verification, repeat (p+q) / 2 times, calculate the average accuracy of (p+q) / 2 iterations, and select the model with the closest average error as the optimal training model to improve the stability and generalization ability of the model, where N is the number of sub-sample sets, p is the number of training sample sets, q is the number of verification sample sets, and the ratio of p to q is 7:3.
[0029] Step 3: Build a Transformer model for fault location. Considering the impact of changes in the distribution network operating status or missing collected data on the accuracy of section positioning, a Transformer-based fault section location model is built with the advantage of the multi-head attention mechanism in capturing the multi-directional, multi-scale, and highly sensitive traveling wave head time series data dependency characteristics.
[0030] a. Distribution network fault section location Transformer model structure is divided into embedding layer, encoder-decoder layer and logistic regression layer. Its structure is as follows Figure 2 shown.
[0031] b. Considering that Transformer cannot capture the relative position relationship between different time series, the time series data must be input into the encoding component in the embedding layer for position encoding in combination with the fault segment. For the task of locating the fault segment, the input of this method is the traveling wave time series matrix constructed by the time when the traveling wave reaches each measurement unit. The matrix order and number of elements are also uncertain, that is, the length of the input is unknown; but for inputs of different lengths, the output length is fixed, which is the fault segment, that is, the fault label in the sample.
[0032] Multi-point wave time series data presents sequential characteristics. The recurrent neural network iteratively processes each time step in the input sequence to give the network model position information, but this method cannot achieve parallel computing and must be operated in series. Since the Transformer cannot capture the relative position relationship between different time points, when the wave head time data is input into the encoding component for encoding, the data column must be positionally encoded (Positional Encoding, PE) to identify the relative position relationship of the input sequence. The position encoding formula is as follows: In formula (2), k is the position index of the traveling wave time series data in the input traveling wave time series data set; d model is the dimension of word embedding of the traveling wave time series data sequence; a can traverse different dimensions of the encoding vector as the calculation progresses; 2a and 2a+1 represent the even and odd positions in the encoding vector respectively; sin and cos functions are used to calculate the encoding values of each position in the even and odd dimensions.
[0033] c. Combined with the relationship between wave head time and fault point, the multi-head attention mechanism is used at the encoder-decoder layer to identify important elements in the time series to assist in prediction. For example, a heavier weight is given to smaller elements and a lower weight is given to larger elements. The calculation steps are as follows: In formula (3), h represents the number of multi-head attention heads; is the weight matrix of the h-th attention head Q, K, V; W is the multi-head attention weight matrix.
[0034] d. The regression layer expands the output signal into a one-dimensional vector after passing it through the fully connected layer, the Softmax layer, and the regularization layer. The corresponding output content is the fault label, which is mapped to the fault section location result. Figure 2 shown.
[0035] Step 4: The characteristics of the model evaluation indicators are as follows: Considering that data-driven distribution network segment positioning is an unbalanced classification task, the confusion matrix is introduced, covering evaluation indicators such as accuracy, recall, precision, and F1 value to comprehensively evaluate the performance of the built model. The definitions of various indicators are shown in formula (4): In formula (4), TP represents the number of true results and model prediction results that are the same, FP represents the number of differences between the predicted results and the true results, and FN represents the number of differences between the true results and the predicted results. c Indicates the prediction accuracy of the model for all data, R ecall It indicates the correct proportion of predicted results based on the actual results, P reision Indicates the proportion of true results in the predicted results. F1 is R ecall and P reision The harmonic mean of can more objectively reflect the model's prediction ability for unstable samples.
[0036] Step 5: Perform model verification based on critical regions. The steps are as follows: a. Considering the case of traveling wave measurement unit n-1, traverse the non-observable area in the distribution network line, and focus on selecting fault samples at the junction of the observable area and the non-observable area as the test set.
[0037] b. Considering the influence of branch circuits on traveling wave transmission, the fault samples at the junction of branch areas are selected as the test set.
[0038] c. The test sample set at the junction is still divided into a fault section every s kilometers, and a fault point is set every t / 2 kilometers in each section.
[0039] Follow the above steps to complete the network division. Figure 3 shown.
[0040] Example 1: Based on a typical 10KV distribution network in a certain area, a distribution network simulation model was built using PSCAD simulation software to generate training data. The network topology diagram is shown in Figure 1 As shown in the figure, the installation positions of the traveling wave acquisition device are nodes A, E, J, L, and N. In order to generate fault data, a fixed length of 1 km is selected to set the fault area in the five lines of bc, de, iI, hH, and lN. The fault time lasts for 0.15 s and the sampling frequency is 10 MHz.
[0041] Due to the complex structure of numerous line branches, in theory, the fault location near the junction of the fault section is prone to misjudgment (because the multi-point traveling wave time series data of the fault at the junction is more similar than the fault point in the middle of the section), that is, the closer the fault location is to the junction, the more difficult it is to locate the fault. Therefore, in order to verify the performance of this model, the training samples and test samples in each fault section are selected with a fixed length of 1km, and the fault points are evenly distributed at the junction of the fault section and in the fault section. The fault section number is shown in Table 1, and the specific location is in Figure 3 Indicated in the middle. Table 1 Sample sampling area
[0042] The above-mentioned fault sections can be selected to represent three fault conditions: trunk line section fault, branch line fault and adjacent line section fault.
[0043] Deep learning must fully explore the rules hidden behind the data to train the model, and a large amount of labeled sample data must be used as support. For large-span distribution networks, this method selects a simulated fault area of 1km. Considering the richness of sample data, this method uses a small step of 50m as the fault point interval to locate the small-span fault section. In order to obtain a sample set of fault data, different fault types, transition resistances, fault initial angles, fault locations, and fault areas are simulated to obtain a simulation parameter traversal table, as shown in Table 2. Table 2 Simulation parameter traversal table
[0044] According to Table 2, 5760 groups of original sample libraries S of fault data are obtained by simulation, and the training set and test set are randomly divided into 4608 groups and 1152 groups according to the ratio of 4:1. The original sample library is divided into N mutually exclusive sub-sample sets of the same size, and N-fold cross-training verification is performed. Then, interference factors are added to the original sample library, and the above training process is repeated.
[0045] During the training of the Transformer model, the number of encoder and decoder layers and the number of heads of the multi-head self-attention mechanism will directly affect the training results. Too many encoder and decoder layers will slow down the model convergence, vanishing gradients, and overfitting, and will increase the complexity of the model. Therefore, in order to avoid the problems of reduced model efficiency and overfitting, it is necessary to experiment and adjust the model during training to find the best combination of parameters to achieve the best training results. This article selects 1, 2, 7 and 14 heads for the experiment, and keeps the other network parameters the same. After 50 EPOCH training, the accuracy of the four schemes is as follows: Figure 4 The evaluation indicators are shown in Table 3. Table 3 Comparison results of evaluation indicators with different numbers of attention heads
[0046] The experimental results show that the prediction accuracy of the model when the number of attention heads is 2 and 7 is better than that when the number of heads is 1 and 14, and the stability of the model when the number of heads is 2 is better than that of the number of heads 7. Considering comprehensively, the number of attention heads of the model is selected as 2. In the Transformer model, the number of encoder-decoder layers is 6. This paper adopts the grid search method to optimize the cases where the number of encoder-decoder layers is 5, 6 and 7. During training, the cross entropy function commonly used in classification problems is selected as the loss function. After 50 EPOCH training, the running time and accuracy of each scheme are shown in Table 4. Table 4 Comparison results of different codec layers
[0047] The results show that when the encoder-decoder layer structure is 5+5, 5+6, 5+7, and 6+5, the accuracy rate exceeds 98%, showing a good prediction effect. Considering the impact of the model running time, the model training performance is best when the layer structure is 5+5. In summary, the training and test set parameters are selected: the number of attention heads is 2, and the number of encoder-decoder layers is 5. This example collects a total of 5760 sets of travel wave time data, of which 80% of the data is used as a training data set and 20% is used for testing. Multi-fold cross-validation is performed according to this ratio, and the SGD optimizer is used to update the parameters. The initial learning rate is set to 0.0001.
[0048] In an ideal sample error-free environment, the Transformer network is used to train the generated data set. The confusion matrix of the network training result is as follows: Figure 5 shown.
[0049] from Figure 5 , Figure 6 As can be seen from Table 5, the accuracy of fault section location by this method can reach 99.4%, which has high precision. Among them, the fault conditions of area 1, area 4, and area 6 can be accurately distinguished, and the positioning accuracy of the remaining areas is relatively high, with only single-digit errors. Misjudgments are mostly concentrated in the boundary areas 2 and 3. Since the samples of adjacent sections are relatively similar, misjudgments are easy to occur, which is consistent with the above theoretical analysis of critical areas. Table 5 Transformer model evaluation results
[0050] Example 2: In the actual operation of the distribution network, as the operating state of the distribution network changes, the topological structure and the traveling wave measurement points will change accordingly. In addition, there may be data pollution in the actual application scenario. In order to verify the ability of attention operation to suppress abnormal features, the model is trained with abnormal sample sets to have a certain generalization ability, thereby improving the accuracy of the model's section positioning under various abnormal conditions. As shown in Table 6, the data samples of the training model in Example 1 are continued to be used, and three abnormal data samples are introduced to simulate the changes in the actual operating state of the distribution network. Table 6 Composition of abnormal training samples
[0051] The fault area determination experiment was compared for the models trained using the three sample input methods: Method 1 is to use only the generated original fault samples S to train the Transformer network model, and use the distribution network simulation model built by PSCAD to generate 2304 sets of samples of real distribution network operation status changes including three types of abnormalities to test the trained model; The second method is to use real abnormal data to expand the sample. According to the three types of abnormal data, 5760 sets of training samples of real distribution network operation status changes are generated in a ratio of 1:1:1, and the original samples and abnormal samples are mixed and recorded as sample S′. The training set and test set are still randomly divided in a ratio of 4:1. Method three is to expand the sample by simulating the composition data, randomly distribute the original fault sample S data, average the data to generate three types of abnormal data, and mix the original sample and the generated sample as sample S″ to participate in the training together, and randomly divide the training set and the test set according to the ratio of 4:1.
[0052] To ensure the accuracy of the comparison results of the three sample input methods, the number of abnormal samples in the new training set generated samples S″ and the real sample S′ is consistent, and the number of test sets is consistent, both of which are 20% of the total sample set, so as to achieve a completely balanced distribution of data set categories. The confusion matrix of the test results of each method and the accuracy of fault section location Figures 7 to 12 And as shown in Table 7. Table 7 Comparison of the accuracy of the three methods
[0053] The test results show that compared with the model tested with the original sample S, after the balanced training model is trained using data augmentation, most abnormal samples can be accurately identified, which effectively improves the generalization ability of the Transformer model. The accuracy of the model trained with abnormal samples has been significantly improved.
[0054] By comparing the second method and the third method, it can be seen that the accuracy of the models trained using the real data augmented sample S′ and the generated data augmented sample S″ differs by only 0.2%, indicating that the three generated data types accurately simulate the real abnormal data characteristics. And no matter whether the generated samples or the real data training samples are used, the multi-head attention mechanism can effectively learn the sample characteristics, the trained model has high performance, strong generalization ability, and high fault location accuracy, indicating that this method has excellent practicality.
Claims
1. A new method for locating distribution network fault sections based on time series data under fault self-healing configuration, characterized by: The method includes a multi-point traveling wave measurement unit configuration structure for fault self-healing, sample construction driven by time series data, Transformer model construction for fault location, model evaluation indicators, and a model verification method based on critical regions.
2. According to claim 1, a new method for locating a distribution network fault section based on time series data driven in a fault self-healing configuration, characterized in that: The configuration structure of the multi-point traveling wave measurement unit for fault self-healing is as follows: accurate positioning of the fault section is the basis of fault self-healing, unified hardware configuration, considering the complexity of the distribution network topology structure, and the configuration requires that the fault in any section of the distribution network meets the observability of the multi-point traveling wave positioning principle.
3. According to claim 1, a new method for locating a distribution network fault section based on time series data driven in a fault self-healing configuration, characterized in that: The sample construction driven by time series data is as follows: The comprehensiveness and generalization of the sample are the basis of the data-driven method. For fault location in the distribution network section, the sample construction steps are as follows: Step 1: Configure the collection devices numbered 1 to n for fault self-healing. set When the judgment is started, the arrival time of the traveling wave head of each configuration unit is collected from t1 to t n Construct a time series sample, where n is the number of configured acquisition devices, U0 is the zero sequence voltage, U set is the zero-sequence voltage threshold for fault determination; Step 2: Divide the distribution network line into a fault section every s kilometers. Considering the positioning accuracy of the fault section, the range of the section is s≤1km. Set a fault point every t kilometers in each section, simulate different fault types, transition resistances, and fault initial angles, and construct a time series sample set. Where s is the length of the fault cell segment and t is the distance between the fault points. Step 3: Considering the time loss, anomalies and changes in network topology during normal operation, the model sample data is replaced with 0 values, that is, some data points in the model sample data are replaced with 0 to simulate the above abnormal conditions, modify the interference factors of error time, expand the sample set, and improve the generalization ability of the model; Step 4: Divide the sample set into N mutually exclusive sub-sample sets of the same size, arbitrarily select p sample sets for training and q sample sets for verification, repeat (p+q) / 2 times, calculate the average accuracy of (p+q) / 2 iterations, and select the model with the closest average error as the optimal training model to improve the stability and generalization ability of the model, where N is the number of sub-sample sets divided, p is the number of sample sets for training, q is the number of sample sets for verification, and the ratio of p to q is 7:
3.
4. According to claim 1, a new method for locating a distribution network fault section based on time series data driven in a fault self-healing configuration, characterized in that: The Transformer model for fault location is constructed as follows: Considering the impact of changes in the distribution network operating status or missing collected data on the accuracy of section positioning, with the advantage of the multi-head attention mechanism in capturing the multi-directional, multi-scale, and highly sensitive traveling wave head time series data dependency characteristics, a Transformer-based fault section location model is constructed: Step 1: Distribution network fault section location Transformer model structure is divided into embedding layer, encoder-decoder layer and logistic regression layer; Step 2: Considering that Transformer cannot capture the relative position relationship between different time series, the time series data is input into the encoding component in the embedding layer and position encoding must be performed in combination with the fault segment. The position encoding formula is as follows: Where: k is the position index of the traveling wave time series data in the input time series data set; d model is the dimension of the word embedding of the traveling wave time series data sequence; a can be used to traverse different dimensions of the encoding vector as the calculation progresses; 2a and 2a+1 represent the even and odd positions in the encoding vector respectively; the sin and cos functions are used to calculate the encoding value of each position in the even and odd dimensions; Step 3: Combined with the relationship between wave head time and fault point, the encoder-decoder layer uses a multi-head attention mechanism to identify important elements in the time series to assist in prediction, assigning heavy weights to small elements and low weights to large elements. The calculation process is as follows: Where: h represents the number of multi-head attention heads; is the weight matrix of the h-th attention head Q, K, V; W is the multi-head attention weight matrix; Step 4: After the output signal passes through the fully connected layer, the Softmax layer, and the regularization layer, the regression layer expands it into a one-dimensional vector. The corresponding output content is the fault label, which is mapped to the fault section location result.
5. According to claim 1, a new method for locating a distribution network fault section based on time series data driven in a fault self-healing configuration, characterized in that: The model evaluation indicators are as follows: Considering that data-driven distribution network segment positioning is an unbalanced classification task, the confusion matrix is introduced to cover evaluation indicators such as accuracy, recall, precision, and F1 value to comprehensively evaluate the performance of the built model: Where: TP represents the number of true results and model prediction results that are the same, FP represents the number of differences between the predicted results and the true results, FN represents the number of differences between the true results and the predicted results, and A c Indicates the prediction accuracy of the model for all data, R ecall It indicates the correct proportion of predicted results based on the actual results, P reision It indicates the proportion of true results in the predicted results, and F1 is R ecall and P reision The harmonic mean of can more objectively reflect the model's prediction ability for unstable samples.
6. According to claim 1, a new method for locating a distribution network fault section based on time series data driven in a fault self-healing configuration, characterized in that: The model validation method based on critical regions is as follows: Step 1: Consider the case of traveling wave measurement unit n-1, traverse the non-observable area in the distribution network line, and focus on selecting fault samples at the junction of the observable area and the non-observable area as the test set; Step 2: Considering the influence of branch circuits on traveling wave transmission, the fault samples at the junction of branch areas are selected as the test set.
Citation Information
Patent Citations
High voltage transmission line traveling wave fault location method considering multi-point faults
CN106093702A
New single-ended fault traveling wave fault location algorithm based on deep learning
CN114966320A
Power transmission line bolt fault detection method based on improved YOLOX
CN117333869A
Full-chain interpretable power distribution network single-phase earth fault type identification method and system
CN117929913A
Cable early fault diagnosis system and method thereof
CN118275822A
Cited By
Power distribution network automation control method and system with fault self-identification function
CN121124366A
