Automatic test method and device for high-voltage flexible direct-current protection equipment based on mean shift, and storage medium

Through mean drift clustering analysis and BP neural network, an automatic testing model for flexible DC protection equipment was established, which solved the problems of low testing efficiency and safety hazards in the existing technology, and achieved efficient and accurate automatic testing.

CN119066457APending Publication Date: 2024-12-03ELECTRIC POWER RESEARCH INSTITUTE OF STATE GRID JIBEI ELECTRIC POWER CO LTD +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411002249.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-25
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

The on-site testing of existing flexible DC protection equipment is inefficient, has high labor costs and safety risks. The program modification is complicated and easy to make it difficult to achieve efficient automatic testing.

Method used

Using a clustering analysis method based on mean drift and backpropagation BP neural network, a prediction model of fault location and transition resistance is established through the clustering and classification of historical test data, automatic testing scheme is optimized, the number of tests is reduced, and the testing efficiency is improved.

Benefits of technology

The test accuracy is achieved by more than 98%, and the number of tests is reduced to less than 50% of the original plan, which improves the testing efficiency and reduces labor costs.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119066457A_ABST
    Figure CN119066457A_ABST
Patent Text Reader

Abstract

The invention relates to an automatic test method and device for high-voltage flexible direct-current protection equipment based on mean shift and a storage medium, and the method comprises three stages of model classification, model prediction and automatic test scheme optimization, and comprises the steps: carrying out the classification of historical data of a protection test through employing an unsupervised machine learning method of mean shift; obtaining a typical scene of a protection device performance test; establishing a neural network prediction model for protecting the test performance; and formulating an optimized protection device automatic test scheme according to the actual measurement information of the typical scene and the test prediction model. According to the method, division of test scenes is completed through a mean shift clustering method, and typical test scene information is provided; through fusion optimization of a test prediction model of a neural network and typical test information, an automatic test scheme of the flexible DC protection device is provided, a simulation result shows that the accuracy of the automatic test method can reach 98% or above, but the test frequency can be reduced to 50% or below of an original test scheme.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of automatic testing of electrical protection devices, relates to the classification and automatic testing of protection tests, and specifically relates to an automatic testing method, device and storage medium for high-voltage flexible DC protection equipment based on mean drift. Background Art

[0002] With the implementation of the national strategy of "carbon peak and carbon neutrality", the advancement of new power system construction and the gradual commissioning of flexible DC projects, the on-site testing of existing flexible DC protection equipment can only protect the export transmission by modifying the internal programs or parameters of the protection device. Repeated and multiple program modifications are inefficient, have high labor costs and pose major safety hazards. Summary of the invention

[0003] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a high-voltage flexible DC protection device automatic testing method, device and storage medium based on mean shift. The method uses a machine learning algorithm to obtain the classification characteristics of flexible DC protection device test scenarios in the context of short test cycles, high labor costs, and cumbersome and error-prone program modifications. The test results are predicted by combining the measured information and historical information of typical scenarios, and an optimized automatic testing plan is formulated to improve test efficiency.

[0004] The present invention solves the technical problem by the following technical solutions:

[0005] A method for automatically testing a high-voltage flexible DC protection device based on mean shift, the method comprising the following steps:

[0006] Obtain protection test history data, including fault location, transition resistance, and protection action status;

[0007] Perform cluster analysis on the acquired historical test data, determine the test scenario classification, and obtain the cluster head information of the class;

[0008] In different test scene classifications, test points are selected and tested in combination with the corresponding cluster head information, and the test results are divided into training sets and test sets;

[0009] The training set is used to train the back propagation BP neural network prediction model with fault location and transition resistance as input and protection action information as output;

[0010] The back propagation BP neural network prediction model is evaluated using the test set. If the test requirements are met, the test is terminated; if the test requirements cannot be met, additional test points are added to repeat the training and evaluation until the test requirements are met.

[0011] Furthermore, the cluster analysis of the acquired historical test data to determine the test scenario classification and obtain the cluster head information of the class is specifically as follows:

[0012] A clustering method based on the mean shift algorithm is used to find clusters along the density increase point. By calculating the position with the largest probability density of the sample point in the feature space, the probability density function is converged to the steady-state point to achieve classification. The formula of the mean shift method is as follows:

[0013]

[0014] Where: M(x) is the offset of the test point within the clustering window;

[0015] x is the cluster center test point, and its value is updated as the drift vector is updated;

[0016] x v is the test point within the bandwidth of the clustering window;

[0017] n is the number of test points in the clustering window;

[0018] is the kernel function, and the Gaussian kernel function is used, and its expression is shown in formula (2):

[0019]

[0020] Given an initial point x, kernel function G(x), and allowable error ε, the mean shift algorithm loops through the following steps until the end condition is met:

[0021] (1) Set the initial value,

[0022] (2) Calculate m according to formulas (1) and (2) v (x i );

[0023] (3) i = i + 1, calculate m v (x i+1 ), if ||m v (x i+1 )-m v (x i+1 )||<ε, end the program; otherwise, jump to (2) to continue calculation.

[0024] Furthermore, the method of selecting test points in different test scene classifications for testing and dividing the test results into a training set and a test set is as follows:

[0025] According to the classification of test scenarios, measurements are performed type by type. Within a certain test class, measurement points are selected according to the 20% average distance from the cluster head position, and actual measurements are performed. The measured data of the measurement points at 10% of the average distance from the cluster head position are used as the training set, and the measured data of the measurement points at 20% of the average distance from the cluster head position are used as the test set.

[0026] Furthermore, the back propagation BP neural network prediction model trained by the training set with the fault location and transition resistance as input and the protection action information as output is specifically:

[0027] For the training data, a BP neural network is used to establish a prediction model with fault location and transition resistance as input and protection action information as output, and this model is used to predict the protection action in the test set. The training of the BP neural network includes two stages: forward propagation of data and back propagation of errors. In the forward propagation stage, data flows into the network from the input layer and is calculated at each layer to obtain the final result of the output layer. The mathematical expression of this process is:

[0028]

[0029] Where: i represents the index of the neuron in the previous layer (layer l-1), j represents the index of the neuron in the current layer (layer l);

[0030] Represent the activation output of the jth neuron in the l-1th layer and the lth layer respectively;

[0031] is the weight connecting the i-th neuron in the l-1th layer and the j-th neuron in the lth layer;

[0032] is the bias of the jth neuron in the lth layer;

[0033] f is the activation function;

[0034] In the back propagation phase, the adjustment of network parameters is to quantify the error between the actual output of the output layer and the expected output, and optimize according to the error, so as to reduce the error and improve the accuracy of the model. The mathematical expression is as follows:

[0035]

[0036] Where: n is the number of output nodes;

[0037] d j is the target value of the jth output neuron of the network;

[0038] y j is the actual output value of the jth output neuron of the network.

[0039] Furthermore, the back propagation BP neural network prediction model is evaluated using the test set. If the test requirements are met, the test is terminated; if the test requirements cannot be met, additional test points are added to repeat the training and evaluation until the test requirements are met. Specifically:

[0040] The measured values ​​of the protection actions in the test set are compared with the predicted values. When the average prediction error is less than a certain value ε, the prediction model is considered accurate and the test results are given using the prediction model; otherwise, measurement points are selected in an advanced manner according to the average distance of 20% from the cluster head position, and actual measurements are performed. The above operations are repeated, and a part of all measured data is used as a training set to establish a neural network prediction model with fault location and transition resistance as input and protection action information as output. All measured data are used as a test set, and the measured values ​​are compared with the predicted values. When the average prediction error is less than a certain value ε, the prediction model is considered accurate and the test results are given. Otherwise, the above operations are repeated until the measurement termination condition is met, that is, the test average error of the test set is less than the fixed value ε.

[0041] An automatic test device for high-voltage flexible DC protection equipment based on mean shift, characterized in that it includes:

[0042] Data acquisition module, used to obtain historical test data of protection test, including fault location, transition resistance, and protection action status;

[0043] The cluster analysis module is used to perform cluster analysis on the acquired historical test data, determine the test scenario classification, and obtain the cluster head information of the class;

[0044] The test module is used to select test points and conduct tests in combination with the corresponding cluster head information in different test scene classifications, and divide the test results into training sets and test sets;

[0045] The model training module uses the training set to train the back propagation BP neural network prediction model with fault location and transition resistance as input and protection action information as output;

[0046] The model testing module uses the test set to evaluate the back propagation BP neural network prediction model. If the test requirements are met, the test is terminated; if the test requirements cannot be met, additional test points are added to repeat the training and evaluation until the test requirements are met.

[0047] Moreover, the cluster analysis module includes a test scenario determination module and a cluster head information acquisition module. The cluster head information acquisition module adopts a clustering method based on the mean shift algorithm to find clusters along the density rise point, and by calculating the position with the largest probability density of the sample point in the feature space, the probability density function converges to the steady-state point to achieve classification. The formula of the mean shift method is as follows:

[0048]

[0049] Where: M(x) is the offset of the test point within the clustering window;

[0050] x is the cluster center test point, and its value is updated as the drift vector is updated;

[0051] x v is the test point within the bandwidth of the clustering window;

[0052] n is the number of test points in the clustering window;

[0053] is the kernel function, and the Gaussian kernel function is used, and its expression is shown in formula (2):

[0054]

[0055] Given an initial point x, kernel function G(x), and allowable error ε, the mean shift algorithm loops through the following steps until the end condition is met:

[0056] (1) Set the initial value,

[0057] (2) Calculate m according to formulas (1) and (2) v (x i );

[0058] (3) i = i + 1, calculate m v (x i+1 ), if ||m v (x i+1 )-m v (x i+1 )||<ε, end the program; otherwise, jump to (2) to continue calculation.

[0059] A computer storage medium having a computer program stored thereon, characterized in that when the computer program is executed by a processor, the steps of the automatic testing method for high-voltage flexible DC protection equipment based on mean shift are implemented.

[0060] The advantages and beneficial effects of the present invention are:

[0061] The present invention uses the mean shift clustering method to cluster test scenarios using historical measurement data, and provides training sets and test sets for prediction using artificial intelligence methods; a prediction model with fault location and transition resistance as input and protection action as output is established using a BP neural network, and an automatic test scheme for a flexible DC protection device is established using the prediction model and actual measurement data. Simulation results show that the accuracy of the automatic test method can reach more than 98%, but the number of tests can be reduced to less than 50% of the original test scheme. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 It is a schematic diagram of the process of the present invention;

[0063] Figure 2 This is a schematic diagram of the PSCAD simulation flexible straight system of the present invention;

[0064] Figure 3 This is a clustering diagram of the present invention with 1980 test samples;

[0065] Figure 4 This is the result diagram of predicting the protection actions of 1980 test points using the measured information of the test points with 20% average distance from the cluster head position;

[0066] Figure 5 The error distribution diagram of predicting the protection action of 1980 test points using the measured information of the test points with 20% average distance from the cluster head position of each cluster;

[0067] Figure 6 This is the result diagram of predicting the protection actions of 1980 test points using the measured information of the test points with a distance of 40% of the average distance from the cluster head position;

[0068] Figure 7 This is a diagram showing the error distribution of predicting the protection actions of 1980 test points using the measured information of test points that are 40% of the average distance from the cluster head position. DETAILED DESCRIPTION

[0069] The present invention is further described in detail below through specific examples. The following examples are only illustrative and not restrictive, and the protection scope of the present invention cannot be limited thereto.

[0070] The present invention uses the unsupervised machine learning method of mean shift to analyze the historical data of protection tests, complete the clustering of typical scenarios of protection device performance testing, obtain the cluster head location information within the class, determine the test points in each class according to the progress of 20% distance from the cluster head, and conduct actual measurements; divide the measured data into a training set and a test set, use the training set to establish a neural network prediction model for protection test performance, use the test set to detect the performance of the prediction model, terminate the test when the average error meets the measurement requirements, and use the prediction model to assign values ​​to all points to be tested. Under the premise of ensuring the test requirements, reduce the number of tests and form an optimized automatic test plan for protection devices;

[0071] like Figure 1 As shown, a high-voltage flexible DC protection device automatic testing method based on mean drift is innovative in that the steps of the method are:

[0072] Obtain protection test history data, including fault location, transition resistance, and protection action status;

[0073] Perform cluster analysis on the acquired historical test data, determine the test scenario classification, and obtain the cluster head information of the class;

[0074] In different test scene classifications, test points are selected and tested in combination with the corresponding cluster head information, and the test results are divided into training sets and test sets;

[0075] The training set is used to train the back propagation BP neural network prediction model with fault location and transition resistance as input and protection action information as output;

[0076] The back propagation BP neural network prediction model is evaluated using the test set. If the test requirements are met, the test is terminated; if the test requirements cannot be met, additional test points are added to repeat the training and evaluation until the test requirements are met.

[0077] The cluster analysis of the acquired historical test data to determine the test scenario classification and obtain the cluster head information of the class is specifically as follows:

[0078] A clustering method based on the mean shift algorithm is used to find clusters along the density increase point. By calculating the position with the largest probability density of the sample point in the feature space, the probability density function is converged to the steady-state point to achieve classification. The formula of the mean shift method is as follows:

[0079]

[0080] Where: M(x) is the offset of the test point within the clustering window;

[0081] x is the cluster center test point, and its value is updated as the drift vector is updated;

[0082] x v is the test point within the bandwidth of the clustering window;

[0083] n is the number of test points in the clustering window;

[0084] is the kernel function, and the Gaussian kernel function is used, and its expression is shown in formula (2):

[0085]

[0086] Given an initial point x, kernel function G(x), and allowable error ε, the mean shift algorithm loops through the following steps until the end condition is met:

[0087] (1) Set the initial value,

[0088] (2) Calculate m according to formulas (1) and (2) v (x i );

[0089] (3) i = i + 1, calculate m v (x i+1 ), if ||m v (xi+1 )-m v (x i+1 )||<ε, end the program; otherwise, jump to (2) to continue calculation.

[0090] The method of selecting test points in different test scene classifications for testing and dividing the test results into a training set and a test set is as follows:

[0091] According to the classification of test scenarios, measurements are performed type by type. Within a certain test class, measurement points are selected according to the 20% average distance from the cluster head position, and actual measurements are performed. The measured data of the measurement points at 10% of the average distance from the cluster head position are used as the training set, and the measured data of the measurement points at 20% of the average distance from the cluster head position are used as the test set.

[0092] The back propagation BP neural network prediction model trained by the training set with the fault location and transition resistance as input and the protection action information as output is specifically:

[0093] For the training data, a BP neural network is used to establish a prediction model with fault location and transition resistance as input and protection action information as output, and this model is used to predict the protection action in the test set. The training of the BP neural network includes two stages: forward propagation of data and back propagation of errors. In the forward propagation stage, data flows into the network from the input layer and is calculated at each layer to obtain the final result of the output layer. The mathematical expression of this process is:

[0094]

[0095] Where: i represents the index of the neuron in the previous layer (layer l-1), j represents the index of the neuron in the current layer (layer l);

[0096] Represent the activation output of the jth neuron in the l-1th layer and the lth layer respectively;

[0097] is the weight connecting the i-th neuron in the l-1th layer and the j-th neuron in the lth layer;

[0098] is the bias of the jth neuron in the lth layer;

[0099] f is the activation function;

[0100] In the back propagation phase, the adjustment of network parameters is to quantify the error between the actual output of the output layer and the expected output, and optimize according to the error, so as to reduce the error and improve the accuracy of the model. The mathematical expression is as follows:

[0101]

[0102] Where: n is the number of output nodes;

[0103] d j is the target value of the jth output neuron of the network;

[0104] y j is the actual output value of the jth output neuron of the network.

[0105] The back propagation BP neural network prediction model is evaluated using the test set. If the test requirements are met, the test is terminated; if the test requirements cannot be met, additional test points are added to repeat the training and evaluation until the test requirements are met. Specifically:

[0106] The measured values ​​of the protection actions in the test set are compared with the predicted values. When the average prediction error is less than a certain value ε, the prediction model is considered accurate and the test results are given using the prediction model; otherwise, measurement points are selected in an advanced manner according to the average distance of 20% from the cluster head position, and actual measurements are performed. The above operations are repeated, and a part of all measured data is used as a training set to establish a neural network prediction model with fault location and transition resistance as input and protection action information as output. All measured data are used as a test set, and the measured values ​​are compared with the predicted values. When the average prediction error is less than a certain value ε, the prediction model is considered accurate and the test results are given. Otherwise, the above operations are repeated until the measurement termination condition is met, that is, the test average error of the test set is less than the fixed value ε.

[0107] An automatic test device for high-voltage flexible DC protection equipment based on mean shift, characterized in that it includes:

[0108] Data acquisition module, used to obtain historical test data of protection test, including fault location, transition resistance, and protection action status;

[0109] The cluster analysis module is used to perform cluster analysis on the acquired historical test data, determine the test scenario classification, and obtain the cluster head information of the class;

[0110] The test module is used to select test points and conduct tests in combination with the corresponding cluster head information in different test scene classifications, and divide the test results into training sets and test sets;

[0111] The model training module uses the training set to train the back propagation BP neural network prediction model with fault location and transition resistance as input and protection action information as output;

[0112] The model testing module uses the test set to evaluate the back propagation BP neural network prediction model. If the test requirements are met, the test is terminated; if the test requirements cannot be met, additional test points are added to repeat the training and evaluation until the test requirements are met.

[0113] The cluster analysis module includes a test scenario determination module and a cluster head information acquisition module. The cluster head information acquisition module adopts a clustering method based on the mean shift algorithm to find clusters along the density rise point, and converges the probability density function to the steady-state point to achieve classification by calculating the position with the largest probability density of the sample point in the feature space. The formula of the mean shift method is as follows:

[0114]

[0115] Where: M(x) is the offset of the test point within the clustering window;

[0116] x is the cluster center test point, and its value is updated as the drift vector is updated;

[0117] x v is the test point within the bandwidth of the clustering window;

[0118] n is the number of test points in the clustering window;

[0119] is the kernel function, and the Gaussian kernel function is used, and its expression is shown in formula (2):

[0120]

[0121] Given an initial point x, kernel function G(x), and allowable error ε, the mean shift algorithm loops through the following steps until the end condition is met:

[0122] (1) Set the initial value,

[0123] (2) Calculate m according to formulas (1) and (2) v (x i );

[0124] (3) i = i + 1, calculate m v (x i+1 ), if ||m v (x i+1 )-m v (x i+1 )||<ε, end the program; otherwise, jump to (2) to continue calculation.

[0125] A computer storage medium having a computer program stored thereon, characterized in that when the computer program is executed by a processor, the steps of the automatic testing method for high-voltage flexible DC protection equipment based on mean shift are implemented.

[0126] Taking the Zhangbei Flexible DC Project as a reference, a four-terminal flexible DC power grid model was built in PSCAD for fault simulation, and this patent was used to evaluate the performance of the automatic testing method for high-voltage flexible DC protection equipment based on mean shift.

[0127] The four-terminal flexible DC grid model is as follows Figure 2 As shown in the figure, the line length of the four-terminal flexible DC grid is 200 km. Each converter station is composed of two modular multilevel converters connected in series and connected using a true bipolar metallic return line. In addition, hybrid DC circuit breakers are configured at both ends of all DC lines to achieve rapid fault isolation after a fault occurs on the DC side.

[0128] This application sets ε=0.03 and conducts automatic test experiments on scenarios with 1980 test points. Based on the fault location, transition resistance, and whether the protection is activated in the relay protection, the test scenario cluster analysis is performed using the mean shift method based on historical test data. Figure 3 The test clustering is shown when the test point is 1980. The algorithm divides the test scene into 8 clusters. Figure 4 In order to select test points within 20% of the cluster average distance from the cluster head position in each cluster for actual measurement, and use the measured data within 10% of the average distance from the cluster head as the training set, and all the measured data as the test set, the BP neural network is used to predict the simulation results of the test, and 997 test data are used to complete the prediction of 1980 test points. Figure 5 For the error distribution, the simulation results show that the prediction accuracy reaches 99.73%, and the actual test times are reduced by 49.65%. Figure 6 , Figure 7 In order to select test points within 40% of the cluster average distance from the cluster head position in each cluster for actual measurement, and use the measured data within 20% of the cluster average distance from the cluster head position as the training set, and all the measured data as the test set, the BP neural network is used to predict the simulation results of the test. Figure 6 The prediction results for 1980 test points are shown using 1097 test data. Figure 7 The simulation results show that the prediction accuracy is 98.54%, and the actual test times are reduced by 44.60%. Both measurements can meet the measurement requirements.

[0129] Although the embodiments and drawings of the present invention are disclosed for illustrative purposes, those skilled in the art will appreciate that various substitutions, changes and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the scope of the present invention is not limited to the contents disclosed in the embodiments and drawings.

Claims

1. An automatic testing method for high-voltage flexible DC protection equipment based on mean shift, characterized in that: The steps of the method are: Obtain protection test history data, including fault location, transition resistance, and protection action status; Perform cluster analysis on the acquired historical test data, determine the test scenario classification, and obtain the cluster head information of the class; In different test scene classifications, test points are selected and tested in combination with the corresponding cluster head information, and the test results are divided into training sets and test sets; The training set is used to train the back propagation BP neural network prediction model with fault location and transition resistance as input and protection action information as output; The back propagation BP neural network prediction model is evaluated using the test set. If the test requirements are met, the test is terminated; if the test requirements cannot be met, additional test points are added to repeat the training and evaluation until the test requirements are met.

2. The automatic testing method for high-voltage flexible DC protection equipment based on mean shift according to claim 1 is characterized in that: The cluster analysis of the acquired historical test data to determine the test scenario classification and obtain the cluster head information of the class is specifically as follows: A clustering method based on the mean shift algorithm is used to find clusters along the density increase point. By calculating the position with the largest probability density of the sample point in the feature space, the probability density function is converged to the steady-state point to achieve classification. The formula of the mean shift method is as follows: Where: M(x) is the offset of the test point within the clustering window; x is the cluster center test point, and its value is updated as the drift vector is updated; x v is the test point within the bandwidth of the clustering window; n is the number of test points in the clustering window; is the kernel function, and the Gaussian kernel function is used, and its expression is shown in formula (2): Given an initial point x, kernel function G(x), and allowable error ε, the mean shift algorithm loops through the following steps until the end condition is met: (1) Set the initial value, (2) Calculate m according to formulas (1) and (2) v (x i ); (3) i = i + 1, calculate m v (x i+1 ), if ||m v (x i+1 )-m v (x i+1 )||<ε, end the program; otherwise, jump to (2) to continue calculation.

3. The automatic testing method for high-voltage flexible DC protection equipment based on mean shift according to claim 1 is characterized in that: The method of selecting test points in different test scene classifications for testing and dividing the test results into a training set and a test set is as follows: According to the classification of test scenarios, measurements are performed type by type. Within a certain test class, measurement points are selected according to the 20% average distance from the cluster head position, and actual measurements are performed. The measured data of the measurement points at 10% of the average distance from the cluster head position are used as the training set, and the measured data of the measurement points at 20% of the average distance from the cluster head position are used as the test set.

4. The automatic testing method for high-voltage flexible DC protection equipment based on mean shift according to claim 1 is characterized in that: The back propagation BP neural network prediction model trained by the training set with the fault location and transition resistance as input and the protection action information as output is specifically: For the training data, a BP neural network is used to establish a prediction model with fault location and transition resistance as input and protection action information as output, and this model is used to predict the protection action in the test set. The training of the BP neural network includes two stages: forward propagation of data and back propagation of errors. In the forward propagation stage, data flows into the network from the input layer and is calculated at each layer to obtain the final result of the output layer. The mathematical expression of this process is: Where: i represents the index of the neuron in the previous layer (layer l-1), j represents the index of the neuron in the current layer (layer l); Represent the activation output of the jth neuron in the l-1th layer and the lth layer respectively; is the weight connecting the i-th neuron in the l-1th layer and the j-th neuron in the lth layer; is the bias of the jth neuron in the lth layer; f is the activation function; In the back propagation phase, the adjustment of network parameters is to quantify the error between the actual output of the output layer and the expected output, and optimize according to the error, so as to reduce the error and improve the accuracy of the model. The mathematical expression is as follows: Where: n is the number of output nodes; d j is the target value of the jth output neuron of the network; y j is the actual output value of the jth output neuron of the network.

5. The automatic testing method for high-voltage flexible DC protection equipment based on mean shift according to claim 1 is characterized in that: The back propagation BP neural network prediction model is evaluated using the test set. If the test requirements are met, the test is terminated; if the test requirements cannot be met, additional test points are added to repeat the training and evaluation until the test requirements are met. Specifically: The measured values ​​of the protection actions in the test set are compared with the predicted values. When the average prediction error is less than a certain value ε, the prediction model is considered accurate and the test results are given using the prediction model; otherwise, measurement points are selected in an advanced manner according to the average distance of 20% from the cluster head position, and actual measurements are performed. The above operations are repeated, and a part of all measured data is used as a training set to establish a neural network prediction model with fault location and transition resistance as input and protection action information as output. All measured data are used as a test set, and the measured values ​​are compared with the predicted values. When the average prediction error is less than a certain value ε, the prediction model is considered accurate and the test results are given. Otherwise, the above operations are repeated until the measurement termination condition is met, that is, the test average error of the test set is less than the fixed value ε.

6. An automatic test device for high-voltage flexible DC protection equipment based on mean shift, characterized in that: include: Data acquisition module, used to obtain historical test data of protection test, including fault location, transition resistance, and protection action status; The cluster analysis module is used to perform cluster analysis on the acquired historical test data, determine the test scenario classification, and obtain the cluster head information of the class; The test module is used to select test points and conduct tests in combination with the corresponding cluster head information in different test scene classifications, and divide the test results into training sets and test sets; The model training module uses the training set to train the back propagation BP neural network prediction model with fault location and transition resistance as input and protection action information as output; The model testing module uses the test set to evaluate the back propagation BP neural network prediction model. If the test requirements are met, the test is terminated; if the test requirements cannot be met, additional test points are added to repeat the training and evaluation until the test requirements are met.

7. The automatic test device for high-voltage flexible DC protection equipment based on mean shift according to claim 6 is characterized in that: The cluster analysis module includes a test scenario determination module and a cluster head information acquisition module. The cluster head information acquisition module adopts a clustering method based on the mean shift algorithm to find clusters along the density rise point, and converges the probability density function to the steady-state point to achieve classification by calculating the position with the largest probability density of the sample point in the feature space. The formula of the mean shift method is as follows: Where: M(x) is the offset of the test point within the clustering window; x is the cluster center test point, and its value is updated as the drift vector is updated; x v is the test point within the bandwidth of the clustering window; n is the number of test points in the clustering window; is the kernel function, and the Gaussian kernel function is used, and its expression is shown in formula (2): Given an initial point x, kernel function G(x), and allowable error ε, the mean shift algorithm loops through the following steps until the end condition is met: (1) Set the initial value, (2) Calculate m according to formulas (1) and (2) v (x i ); (3) i = i + 1, calculate m v (x i+1 ), if ||m v (x i+1 )-m v (x i+1 )||<ε, end the program; otherwise, jump to (2) to continue calculation.

8. A computer storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the automatic testing method of high-voltage flexible DC protection equipment based on mean shift as described in any one of claims 1 to 5 are implemented.