Protection evaluation method and device for protection action boundary value of flexible direct current system based on multi-layer perceptron (MLP), and storage medium
Through the method based on the multi-layer perceptron, an evaluation model of the protection action boundary value of the high-voltage DC system was established, which solved the reliability problem of the protection evaluation of high-voltage DC lines in the existing technology, achieved rapid and accurate protection performance evaluation, and improved power System security and stability.
Patent Information
- Application Number
- CN202410973285.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2024-07-05
- Filing Date
- 2024-07-19
- Publication Date
- 2024-11-22
AI Technical Summary
The prior art has low reliability in high voltage DC line protection evaluation, and it is impossible to accurately and quickly evaluate protection action behaviors, especially in the event of a failure, manual evaluation takes a long time and is inefficient, and cannot adapt to the rapidity requirements of online evaluation.
Using a multi-layer perceptron (MLP)-based method, a classification model and regression model are established by acquiring and cleaning the training set and test set data, predicting the protection action boundary value and fault position boundary value, and then evaluating protection performance.
It realizes a fast and accurate evaluation of the protection action behavior of high-voltage DC system, simplifies the judgment process, provides effective data support for protection performance, and improves the safe and stable operation of the power system.
Smart Images

Figure CN119004166A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of electrical relay protection, and relates to the problem of predicting the boundary value of protection action performance of a flexible direct current system and evaluating the protection performance, and specifically relates to a protection evaluation method, device and storage medium for the boundary value of protection action of a flexible direct current system based on a multi-layer perceptron (MLP). Background Art
[0002] The reliability of the HVDC system control and protection device is crucial to the safe and stable operation of the power system. At present, the reliability of HVDC line protection is low. Accurate and rapid evaluation of the action behavior of HVDC line protection is of great significance to improving the safe and stable operation of the power system.
[0003] At present, most of the research on the evaluation of relay protection behavior at home and abroad focuses on the evaluation of AC protection action behavior, while there are few reports on the situation when the high-voltage DC line fails. In the evaluation research of AC relay protection action, the manual evaluation method is time-consuming and has a great impact on the operating personnel. There is also an exploration of using expert systems for protection evaluation. This type of method has shortcomings such as being limited by knowledge domains, and has low search efficiency and slow speed, and cannot meet the requirements of rapid online evaluation. Summary of the invention
[0004] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a protection evaluation method, device and storage medium for the protection action boundary value of a flexible DC system based on a multi-layer perceptron MLP, which can predict the position interval and transition resistance boundary value of the flexible DC protection, and evaluate the protection performance accordingly. The transition resistance boundary value and the fault position boundary value of the protection action boundary are predicted through limited data of the protection test, providing a basic basis for the evaluation of relay protection function and performance.
[0005] The present invention solves the technical problem by the following technical solutions:
[0006] A protection evaluation method for a flexible DC system protection action boundary value based on a multi-layer perceptron MLP, the method comprising the following steps:
[0007] Get the total training set and test set;
[0008] Perform data cleaning on the total training set and test set data to clearly define the boundaries of the protection actions;
[0009] The MLP classification model is used to establish a prediction model for protection action behavior, and the discrete protection action behavior results are replaced by continuously changing protection action probability values to achieve a fitable result of protection action behavior.
[0010] Based on the fitting result of protection action behavior, the regression model of fault information and protection action is established by using MLP regression model, and the nonlinear function relationship between fault information and protection action behavior is established, so as to solve the protection action boundary.
[0011] Calculate the transition resistance boundary value and the fault position boundary value of the protection action boundary according to the regression model of the fault information and the protection action;
[0012] The protection action performance is evaluated based on the transition resistance boundary value and the fault location boundary value of the protection action boundary.
[0013] Moreover, the acquisition of the total training set and the test set is specifically as follows:
[0014] The flexible DC system uses the field measured or simulated test data of the protection action of the traveling wave protection device when the fault is at different transition resistance values and different fault locations as the training set and test set. The data includes the location information of the fault, the transition resistance information and the protection action behavior information.
[0015] In the MLP classification model, the fault location information and transition resistance value are used as the input layer of the neural network, and the probability of protection action determined by the MLP classification model is used as the output layer;
[0016] In the MLP regression model, the fault location, transition resistance value, and the judgment probability of the protection action by the MLP classification model are used as the neural network input layer, and the predicted transition resistance boundary value and fault location boundary value of the protection action boundary are used as the output layer;
[0017] The mathematical model of MLP input and output is expressed as:
[0018] z i =R(W i z i-1 +b i ) (1)
[0019] Where: z i is the output of the i-th layer;
[0020] W i and b i are the weight and bias between layer i and layer i-1 respectively;
[0021] R is the activation function, which uses the ReLU function;
[0022] The ReLU function is:
[0023] f(x)=max(0,x) (2)
[0024] x is the input vector of the previous layer of neural network input to the neurons in this layer. The neurons using the ReLU function will output according to their position in the network.
[0025] max(0,w T x+b)(3)
[0026] To the next layer of neurons or as the output of the entire neural network.
[0027] Furthermore, the data of the total training set and test set are cleaned to clarify the protection action boundaries:
[0028] Assume that there are m test data in the data set, and there are p fault locations, which increase at equal intervals; each fault location has q test data, and in the q test data, the transition resistance increases at equal intervals; suppose that the protection action state corresponding to each fault location and the transition resistance value is A kj To protect the action, A kj When the value is 0, it means the protection is not in action. kj When the value is 1, it indicates protection action;
[0029] The protection action state matrix is constructed as follows:
[0030]
[0031] In order to check the protection action results under different transition resistance values at the same fault location, a data window M with a horizontal length of 3 and a vertical width of 1 is constructed; in order to check the protection action results under different fault locations with the same transition resistance value, a data window N with a horizontal width of 1 and a vertical length of 2 is constructed; the state value sequence in the data windows M and N will be used as the basis for correcting the protection refusal and misoperation data; let X1 and X2 be the original sequences in the data windows M and N during the inspection process, respectively, and Y is the correct sequence after correction according to the rules;
[0032] Clean the historical test data. The data cleaning rules are as follows:
[0033] Rule 1: M starts from the first row and slides from left to right in each row of the matrix. The sliding step is 1. The three state values from left to right in the data window M after each sliding are recorded to form a state value sequence. If the state value sequence in the data window M is It is considered that the j-2 transition resistance value of the k-th fault position has failed to operate due to the problem of the protection itself, and the state value sequence in the modified data window M is The correction formula is:
[0034]
[0035] Rule 2: M starts from the first row and slides from left to right in each row of the matrix. The sliding step is 1. After each sliding, the three state values of M from left to right in the data window are recorded to form a state value sequence. If the state value sequence in the data window M is Then start the status window N; The data at the bottom of the state window N records the state value sequence composed of the two state values from top to bottom in the data window N after each slide. If the state value sequence of data window N appears It is considered that the j-1 transition resistance value of the k-th fault position has malfunctioned due to the problem of the protection itself, and the state value sequence in the corrected data window M is If the state value sequence of data window N appears It is considered that the j-2 transition resistance value of the k-th fault position fails to operate due to the problem of the protection itself, and the state value sequence in the modified data window M is The correction formula is:
[0036]
[0037] Rule 3: N starts from the first column and slides from top to bottom in each column of the matrix. The sliding step is 1. The data window N is recorded from the top to the bottom to form a state value sequence. If the state value sequence in the data window N is It is considered that the jth transition resistance value of the kth fault position has malfunctioned due to the problem of the protection itself, and the state value sequence in the correction window data N is The correction formula is:
[0038]
[0039] By default, all protection actions in the first fault position do not malfunction.
[0040] Moreover, the prediction model of the protection action behavior is established by using the MLP classification model, and the discrete protection action behavior results are replaced by the continuously changing protection action probability values, so as to achieve the goal that the protection action behavior can be fitted. Specifically,
[0041] The transition resistance value and fault location of the fault setting during the test are used as input, and the protection action result is used as output to train the MLP classification model using formula (1);
[0042] The probability value of the protection action under this fault determined by the MLP classification model is used as the output of the MLP classification model;
[0043] A certain threshold is set, and the output of the MLP classification model is considered as a protection action when it is greater than the threshold, and the output is considered as a protection action when it is less than the threshold. The prediction accuracy of the MLP classification model is calculated by comparing it with the tested protection action behavior.
[0044] The MLP classification model determines the protection action behavior formula:
[0045]
[0046] Among them: act is the protection action behavior judged by the MLP classification model, 1 indicates protection action, and 0 indicates no protection action;
[0047] y_pred_prob is the probability value when the MLP classification model determines that the protection action behavior is 1; threshold is the threshold set in advance;
[0048] The accuracy calculation formula of the MLP classification model is:
[0049]
[0050] The prediction model of the protection action behavior outputs the probability value of the protection action, and the protection action probability is a continuous value protection action behavior between 0 and 1.
[0051] Moreover, the MLP regression model is used to establish a regression model of fault information and protection action, and a nonlinear functional relationship between fault information and protection action behavior is established, and then the protection action boundary is solved as follows:
[0052] The protection action probability value, fault location, and transition resistance value are used as inputs, and the transition resistance boundary value of the protection action boundary and the fault location boundary value are used as outputs to train the MLP regression model. To optimize the training effect, the data is cut near the determined value of the protection action probability value, and only a shorter length of data segment is retained for each fault location as the MLP regression model training set;
[0053] The protection action state matrix after cutting is:
[0054]
[0055] Where: l is the length of the data segment reserved for each fault location.
[0056] Moreover, the calculation of the transition resistance boundary value and the fault position boundary value of the protection action boundary based on the regression model of fault information and protection action is specifically as follows: when the protection action probability value is determined to be a set value and is at the protection action boundary, the fault position boundary value corresponding to a certain transition resistance under the protection action boundary is calculated using the regression model of fault information and protection action; and the transition resistance boundary value corresponding to a certain fault position under the protection action boundary is calculated.
[0057] A protection evaluation device for a flexible DC system protection action boundary value based on a multi-layer perceptron MLP, comprising:
[0058] Data collection module, used to obtain the total training set and test set;
[0059] The data preprocessing module is used to clean the total training set and test set data and clarify the protection action boundaries;
[0060] MLP classification model module, which uses the MLP classification model to establish a prediction model for protection action behavior, and replaces discrete protection action behavior results with continuously changing protection action probability values to achieve a fitable result for protection action behavior;
[0061] The MLP regression model module uses the MLP regression model to establish a regression model of fault information and protection action based on the results of the protection action behavior that can be fitted, and establishes a nonlinear functional relationship between fault information and protection action behavior, thereby solving the protection action boundary; the transition resistance boundary value and the fault position boundary value of the protection action boundary are calculated according to the regression model of fault information and protection action;
[0062] The data output module evaluates the protection action performance according to the transition resistance boundary value and the fault position boundary value of the protection action boundary.
[0063] Moreover, the MLP classification model module includes a regression model building module and a boundary value calculation module;
[0064] The regression model building module uses the protection action probability value, fault location, and transition resistance value as input, and the transition resistance boundary value and fault location boundary value as output to train the MLP regression model. To optimize the training effect, the data is cut near the determined value of the protection action probability value, and only a shorter length of data segment is retained for each fault location as the MLP regression model training set;
[0065] The protection action state matrix after cutting is:
[0066]
[0067] Where: l is the length of the data segment reserved for each fault location;
[0068] The boundary value calculation module calculates the transition resistance boundary value and the fault position boundary value of the protection action boundary, determines that the protection is at the protection action boundary when the protection action probability value is the set value, and uses the regression model of fault information and protection action to calculate the fault position boundary value corresponding to a certain transition resistance under the protection action boundary; calculates the transition resistance boundary value corresponding to a certain fault position under the protection action boundary.
[0069] 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 protection evaluation method for the protection action boundary value of a flexible DC system based on a multi-layer perceptron MLP are implemented.
[0070] The advantages and beneficial effects of the present invention are:
[0071] 1. The present invention directly establishes an association relationship between the transition resistance value, fault location and corresponding protection action behavior through the MLP classification model, without the need to understand complex protection algorithms and logic, thereby simplifying the process of determining whether the flexible DC protection is to be operated;
[0072] 2. The present invention provides the functional relationship between the fault transition resistance value, fault location and protection action probability value through the MLP regression model, providing accurate data support for the evaluation of flexible DC protection results. BRIEF DESCRIPTION OF THE DRAWINGS
[0073] Figure 1 It is a schematic diagram of the process of the present invention;
[0074] Figure 2 It is the basic structure diagram of MLP of the present invention;
[0075] Figure 3 This is a schematic diagram of a four-terminal flexible DC power grid used in the present invention;
[0076] Figure 4 A three-dimensional graph of the data acquired for the test of the present invention;
[0077] Figure 5 This is a three-dimensional graph of the data after cleaning in the present invention;
[0078] Figure 6 This is a schematic diagram of the effect of the MLP classification model of the present invention;
[0079] Figure 7 This is a data graph after data cutting of the present invention;
[0080] Figure 8 This is a comparison diagram of the prediction results of the present invention. DETAILED DESCRIPTION
[0081] 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.
[0082] A protection evaluation method for flexible DC system protection action boundary value based on multi-layer perceptron MLP is innovative in that: Figure 2As shown, the MLP of the present invention includes an input layer, a hidden layer and an output layer, each layer is composed of a plurality of neurons, and each neuron is fully connected by connecting to all neurons of the previous layer. The method adopts an MLP classification model to imitate an actual protection device, and obtains a high-accuracy functional relationship between the protection action probability values under different transition resistances and fault positions through an MLP regression model and a trained MLP classification model. According to this functional relationship, the transition resistance boundary value corresponding to the protection action boundary and the boundary value of the fault position are obtained at a given accuracy, and the boundary value is used as a basis for evaluating the protection tolerance to transition resistance and the protection range;
[0083] The present invention takes into account the complexity of relay protection principles and the difficulty of obtaining a direct description of the relationship between the measured quantity and the protection performance using a simple test method. The present invention uses historical measurement information and a multilayer perceptron (MLP) of a feedforward neural network model to establish a relationship model between fault information and protection action behavior to infer the performance of the protection device.
[0084] like Figure 1 As shown, the steps of the method are:
[0085] Get the total training set and test set;
[0086] Perform data cleaning on the total training set and test set data to clearly define the boundaries of the protection actions;
[0087] The MLP classification model is used to establish a prediction model for protection action behavior, and the discrete protection action behavior results are replaced by continuously changing protection action probability values to achieve a fitable result of protection action behavior.
[0088] Based on the fitting result of protection action behavior, the regression model of fault information and protection action is established by using MLP regression model, and the nonlinear function relationship between fault information and protection action behavior is established, so as to solve the protection action boundary.
[0089] Calculate the transition resistance boundary value and the fault position boundary value of the protection action boundary according to the regression model of the fault information and the protection action;
[0090] The protection action performance is evaluated based on the transition resistance boundary value and the fault location boundary value of the protection action boundary.
[0091] The specific steps of obtaining the total training set and test set are as follows:
[0092] The flexible DC system uses the field measured or simulated test data of the protection action of the traveling wave protection device when the fault is at different transition resistance values and different fault locations as the training set and test set. The data includes the location information of the fault, the transition resistance information and the protection action behavior information.
[0093] The MLP neural network is the basic architecture of the MLP classification model and the MLP regression model. In the MLP classification model, the fault location information and the transition resistance value are used as the input layer of the neural network, and the probability of the protection action determined by the MLP classification model is used as the output layer;
[0094] In the MLP regression model, the fault location, transition resistance value, and the judgment probability of the protection action by the MLP classification model are used as the neural network input layer, and the predicted transition resistance boundary value and fault location boundary value are used as the output layer;
[0095] The mathematical model of MLP input and output is expressed as:
[0096] z i =R(W i z i-1 +b i ) i =R(W i z i-1 +b i ) (1)
[0097] Where: z i is the output of the i-th layer;
[0098] W i and b i are the weight and bias between layer i and layer i-1 respectively;
[0099] R is the activation function, which uses the ReLU function;
[0100] The ReLU function is:
[0101] f(x)=max(0,x) (2)
[0102] x is the input vector of the previous layer of neural network input to the neurons in this layer. The neurons using the ReLU function will output according to their position in the network.
[0103] max(0,w T x+b)(3)
[0104] To the next layer of neurons or as the output of the entire neural network.
[0105] The data cleaning of the total training set and test set data obtained to eliminate unclear protection action demarcation points caused by protection refusal or misoperation is specifically as follows:
[0106] Assume that there are m test data in the data set, and there are p fault locations, which increase at equal intervals; each fault location has q test data, and in the q test data, the transition resistance increases at equal intervals; suppose that the protection action state corresponding to each fault location and the transition resistance value is A kj To protect the action, A kj When the value is 0, it means the protection is not in action. kj When the value is 1, it indicates protection action;
[0107] The protection action state matrix is constructed as follows:
[0108]
[0109] In order to check the protection action results under different transition resistance values at the same fault location, a data window M with a horizontal length of 3 and a vertical width of 1 is constructed; in order to check the protection action results under different fault locations with the same transition resistance value, a data window N with a horizontal width of 1 and a vertical length of 2 is constructed; the state value sequence in the data windows M and N will be used as the basis for correcting the protection refusal and misoperation data; let X1 and X2 be the original sequences in the data windows M and N during the inspection process, respectively, and Y is the correct sequence after correction according to the rules;
[0110] Clean the historical test data. The data cleaning rules are as follows:
[0111] Rule 1: M starts from the first row and slides from left to right in each row of the matrix. The sliding step is 1. The three state values from left to right in the data window M after each sliding are recorded to form a state value sequence. If the state value sequence in the data window M is It is considered that the j-2 transition resistance value of the k-th fault position has failed to operate due to the problem of the protection itself, and the state value sequence in the modified data window M is The correction formula is:
[0112]
[0113] Rule 2: M starts from the first row and slides from left to right in each row of the matrix. The sliding step is 1. After each sliding, the three state values of M from left to right in the data window are recorded to form a state value sequence. If the state value sequence in the data window M is Then start the status window N; The data at the bottom of the state window N records the state value sequence composed of the two state values from top to bottom in the data window N after each slide. If the state value sequence of data window N appears It is considered that the j-1 transition resistance value of the k-th fault position has malfunctioned due to the problem of the protection itself, and the state value sequence in the corrected data window M is If the state value sequence of data window N appears It is considered that the j-2 transition resistance value of the k-th fault position fails to operate due to the problem of the protection itself, and the state value sequence in the modified data window M is The correction formula is:
[0114]
[0115] Rule 3: N starts from the first column and slides from top to bottom in each column of the matrix. The sliding step is 1. The data window N is recorded from the top to the bottom to form a state value sequence. If the state value sequence in the data window N is It is considered that the jth transition resistance value of the kth fault position has malfunctioned due to the problem of the protection itself, and the state value sequence in the correction window data N is The correction formula is:
[0116]
[0117] By default, all protection actions in the first fault position do not malfunction.
[0118] The prediction model of protection action behavior established by using the MLP classification model is specifically:
[0119] The transition resistance value and fault location of the fault setting during the test are used as input, and the protection action result is used as output to train the MLP classification model using formula (1);
[0120] The probability value of the protection action under this fault determined by the MLP classification model is used as the output of the MLP classification model;
[0121] A certain threshold is set, and the output of the MLP classification model is considered as a protection action when it is greater than the threshold, and the output is considered as a protection action when it is less than the threshold. The prediction accuracy of the MLP classification model is calculated by comparing it with the tested protection action behavior.
[0122] The MLP classification model determines the protection action behavior formula:
[0123]
[0124] Among them: act is the protection action behavior judged by the MLP classification model, 1 indicates protection action, and 0 indicates no protection action;
[0125] y_pred_prob is the probability value when the MLP classification model determines that the protection action behavior is 1; threshold is the threshold set in advance;
[0126] The accuracy calculation formula of the MLP classification model is:
[0127]
[0128] The prediction model of protection action behavior outputs the probability value of protection action, and the probability of protection action is a continuous value between 0 and 1. This solves the problem that the MLP regression model cannot be used to establish a regression model of fault information and protection action because the original protection action behavior is only a discrete value of 0 and 1.
[0129] The regression model of fault information and protection action established by using the MLP regression model is specifically:
[0130] The protection action probability value, fault location, and transition resistance value are used as input, and the transition resistance boundary value and fault location boundary value are used as output to train the MLP regression model. To optimize the training effect, the data is cut near the determined value of the protection action probability value, and only a shorter length of data segment is retained for each fault location as the MLP regression model training set;
[0131] The protection action state matrix after cutting is:
[0132]
[0133] Where: l is the length of the data segment reserved for each fault location.
[0134] The regression model of fault information and protection action fits the functional relationship between the two, providing a basis for the subsequent calculation of the protection action boundary value.
[0135] The method of calculating the transition resistance boundary value and the fault position boundary value based on the regression model of fault information and protection action is specifically as follows: when it is determined that the protection action probability value is a set value, the protection is in a boundary action state, and the fault position boundary value corresponding to a certain transition resistance under the protection boundary action is calculated using the regression model of fault information and protection action; and the transition resistance boundary value corresponding to a certain fault position under the protection action boundary is calculated.
[0136] A protection evaluation device for a flexible DC system protection action boundary value based on a multi-layer perceptron MLP, characterized by comprising:
[0137] Data collection module, used to obtain the total training set and test set;
[0138] The data preprocessing module is used to clean the total training set and test set data and clarify the protection action boundaries;
[0139] MLP classification model module, which uses the MLP classification model to establish a prediction model for protection action behavior, and replaces discrete protection action behavior results with continuously changing protection action probability values to achieve a fitable result for protection action behavior;
[0140] The MLP regression model module uses the MLP regression model to establish a regression model of fault information and protection action based on the results of the protection action behavior that can be fitted, and establishes a nonlinear functional relationship between fault information and protection action behavior, thereby solving the protection action boundary; the transition resistance boundary value and the fault position boundary value of the protection action boundary are calculated according to the regression model of fault information and protection action;
[0141] The data output module evaluates the protection action performance according to the transition resistance boundary value and the fault position boundary value of the protection action boundary.
[0142] Read data from the protection test results of the flexible DC system;
[0143] Pre-process the collected data to clarify the demarcation point of the protection action;
[0144] The preprocessed data is processed using the trained MLP classification model to calculate the probability value of the protection action of the flexible DC system;
[0145] The trained MLP regression model is used to process the preprocessed data and the protection action probability value output by the prediction model of the protection action behavior, and a regression model of fault information and protection action is established to calculate the protection action boundary value;
[0146] The regression model of fault information and protection action is used to calculate the transition resistance boundary value and the fault location boundary value under the protection action boundary to form an evaluation of the protection.
[0147] 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 protection evaluation method for the protection action boundary value of a flexible DC system based on a multi-layer perceptron MLP are implemented.
[0148] like Figure 3 As shown, the present invention adopts a four-terminal flexible direct current model system to simulate line faults. The protection principle is traveling wave protection. Whether the output of the system recording meets the traveling wave protection criterion is the standard for judging whether the protection is activated during one operation. The system recording data is used as the training set and the test set, and the MLP classification model and the MLP regression model are trained after data cleaning. During the test, the fault location changes from 1% to 99% of the total length of the line at 1% intervals; the transition resistance value changes from 5 to 100Ω at 5Ω intervals.
[0149] The test data obtained is as follows Figure 4 As shown in the figure, the data after cleaning is as follows Figure 5 After data cleaning, the data of protection refusal and misoperation are corrected.
[0150] The cleaned data is used to train the MLP classification model, and the output is the model's predicted value for the probability of protection action. This operation improves the protection action behavior from a two-value representation of 0 or 1 to a representation of a continuously changing value between 0 and 1. The effect is as follows Figure 6 shown.
[0151] In each fault location, the first value with the model predicted protection action probability value just greater than 0.5 is taken as the data center, and the test data is cut into pieces with a length of 5. If the test parameter boundary value is reached, the length of the data segment is shortened accordingly. The data segments after cutting are as follows: Figure 7 As shown in Figure 2, the cut removes data near non-protected action boundaries that may adversely affect the training of the MLP regression model.
[0152] The MLP regression model is trained using the cut data segment, with the fault location and protection action probability as independent variables and the transition resistance value as the dependent variable. It is determined that when the protection action probability value is 0.5, the protection is in the state of just action. The training results can achieve the prediction of the transition resistance boundary value of the protection action boundary with an accuracy of 1Ω. The results are as follows Figure 7 shown.
[0153] The effect of using the present invention to predict the boundary value is compared with the effect when the model adopts different parameters. Figure 8 As shown in the figure, "transition resistance boundary value -1" and "transition resistance boundary value -0" are the range of transition resistance boundary values when only test data is used to determine the transition resistance boundary value; "predicted transition resistance boundary value -1", "predicted transition resistance boundary value -2", and "predicted transition resistance boundary value -3" are respectively, when the protection action probability value is determined to be 0.4, 0.5, and 0.6, the protection is in the state of just acting, and the boundary resistance prediction value results obtained by the training model. It can be seen from the figure that "predicted transition resistance boundary value -2" has the best fitting effect. Table 1 is an example of the prediction result data of the present invention.
[0154] Table 1
[0155]
[0156] Before the fault position 0.27, the transition resistance boundary value is greater than or equal to 100Ω, which is beyond the test data range. The fault position is the distance protection setting where the fault occurs, accounting for the proportion of the total length of the line; the transition resistance boundary value unit is Ω.
[0157] The present invention realizes the establishment of an MLP model between fault information and protection action behavior through discrete test data, obtains the transition resistance boundary range, saves test time, and provides a basis for protection evaluation. The evaluation of the protection by the calculation example result is: when the power system is running, when the fault occurs within 27% of the total length of the line from the protection installation, the protection tolerance transition resistance value is greater than 99Ω; when the fault occurs at 62% of the total length of the line, the protection tolerance transition resistance value is 50Ω; when the fault occurs at 99% of the total length of the line, the protection tolerance transition resistance value is 24Ω.
[0158] 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. A protection evaluation method for flexible DC system protection action boundary value based on multi-layer perceptron MLP, characterized by: The steps of the method are: Get the total training set and test set; Perform data cleaning on the total training set and test set data to clearly define the boundaries of the protection actions; The MLP classification model is used to establish a prediction model for protection action behavior, and the discrete protection action behavior results are replaced by continuously changing protection action probability values to achieve a fitable result of protection action behavior. Based on the fitting result of protection action behavior, the regression model of fault information and protection action is established by using MLP regression model, and the nonlinear function relationship between fault information and protection action behavior is established, so as to solve the protection action boundary. Calculate the transition resistance boundary value and the fault position boundary value of the protection action boundary according to the regression model of the fault information and the protection action; The protection action performance is evaluated based on the transition resistance boundary value and the fault location boundary value of the protection action boundary.
2. The protection evaluation method for the protection action boundary value of the flexible DC system based on the multi-layer perceptron MLP according to claim 1 is characterized in that: The specific steps of obtaining the total training set and test set are as follows: The flexible DC system uses the field measured or simulated test data of the protection action of the traveling wave protection device when the fault is at different transition resistance values and different fault locations as the training set and test set. The data includes the location information of the fault, the transition resistance information and the protection action behavior information. In the MLP classification model, the fault location information and transition resistance value are used as the input layer of the neural network, and the probability of protection action determined by the MLP classification model is used as the output layer; In the MLP regression model, the fault location, transition resistance value, and the judgment probability of the protection action by the MLP classification model are used as the neural network input layer, and the predicted transition resistance boundary value and fault location boundary value of the protection action boundary are used as the output layer; The mathematical model of MLP input and output is expressed as: z i =R(W i z i-1 +b i ) (1) Where: z i is the output of the i-th layer; W i and b i are the weight and bias between layer i and layer i-1 respectively; R is the activation function, which uses the ReLU function; The ReLU function is: f(x)=max(0,x) (2) x is the input vector of the previous layer of neural network input to the neurons in this layer. The neurons using the ReLU function will output according to their position in the network. max(0,w T x+b)(3) to the next layer of neurons or as the output of the entire neural network.
3. The protection evaluation method for the protection action boundary value of the flexible DC system based on the multi-layer perceptron MLP according to claim 1 is characterized in that: The data cleaning of the total training set and test set data obtained to clarify the protection action boundaries is specifically as follows: Assume that there are m test data in the data set, and there are p fault locations, which increase at equal intervals; each fault location has q test data, and in the q test data, the transition resistance increases at equal intervals; suppose that the protection action state corresponding to each fault location and the transition resistance value is A kj To protect the action, A kj When the value is 0, it means the protection is not in action. kj When the value is 1, it indicates protection action; The protection action state matrix is constructed as follows: In order to check the protection action results under different transition resistance values at the same fault location, a data window M with a horizontal length of 3 and a vertical width of 1 is constructed; in order to check the protection action results under different fault locations with the same transition resistance value, a data window N with a horizontal width of 1 and a vertical length of 2 is constructed; the state value sequence in the data windows M and N will be used as the basis for correcting the protection refusal and misoperation data; let X1 and X2 be the original sequences in the data windows M and N during the inspection process, respectively, and Y is the correct sequence after correction according to the rules; Clean the historical test data. The data cleaning rules are as follows: Rule 1: M starts from the first row and slides from left to right in each row of the matrix. The sliding step is 1. The three state values from left to right in the data window M after each sliding are recorded to form a state value sequence. If the state value sequence in the data window M is It is considered that the j-2 transition resistance value of the k-th fault position has failed to operate due to the problem of the protection itself, and the state value sequence in the modified data window M is The correction formula is: Rule 2: M starts from the first row and slides from left to right in each row of the matrix. The sliding step is 1. After each sliding, the three state values of M from left to right in the data window are recorded to form a state value sequence. If the state value sequence in the data window M is Then start the status window N; The data at the bottom of the state window N records the state value sequence composed of the two state values from top to bottom in the data window N after each slide. If the state value sequence of data window N appears It is considered that the j-1 transition resistance value of the k-th fault position has malfunctioned due to the problem of the protection itself, and the state value sequence in the corrected data window M is If the state value sequence of data window N appears It is considered that the j-2 transition resistance value of the k-th fault position fails to operate due to the problem of the protection itself, and the state value sequence in the modified data window M is The correction formula is: Rule 3: N starts from the first column and slides from top to bottom in each column of the matrix. The sliding step is 1. The state value sequence of the recorded data window N consists of two state values from top to bottom. If the state value sequence in the data window N is It is considered that the jth transition resistance value of the kth fault position has malfunctioned due to the problem of the protection itself, and the state value sequence in the correction window data N is The correction formula is: By default, all protection actions in the first fault position do not malfunction.
4. The protection evaluation method for the protection action boundary value of the flexible DC system based on the multi-layer perceptron MLP according to claim 1 is characterized in that: The prediction model of protection action behavior is established by using the MLP classification model, and the discrete protection action behavior results are replaced by continuously changing protection action probability values to achieve the goal of fitting the protection action behavior. Specifically, The transition resistance value and fault location of the fault setting during the test are used as input, and the protection action result is used as output to train the MLP classification model using formula (1); The probability value of the protection action under this fault determined by the MLP classification model is used as the output of the MLP classification model; A certain threshold is set, and the output of the MLP classification model is considered as a protection action when it is greater than the threshold, and the output is considered as a protection action when it is less than the threshold. The prediction accuracy of the MLP classification model is calculated by comparing it with the tested protection action behavior. The MLP classification model determines the protection action behavior formula: Among them: act is the protection action behavior judged by the MLP classification model, 1 indicates protection action, and 0 indicates no protection action; y_pred_prob is the probability value when the MLP classification model determines that the protection action behavior is 1; threshold is the threshold set in advance; The accuracy calculation formula of the MLP classification model is: The prediction model of the protection action behavior outputs the probability value of the protection action, and the protection action probability is a continuous value protection action behavior between 0 and 1.
5. The protection evaluation method for the protection action boundary value of the flexible DC system based on the multi-layer perceptron MLP according to claim 1 is characterized in that: The MLP regression model is used to establish a regression model of fault information and protection action, to establish a nonlinear functional relationship between fault information and protection action behavior, and then to solve the protection action boundary as follows: The protection action probability value, fault location, and transition resistance value are used as inputs, and the transition resistance boundary value of the protection action boundary and the fault location boundary value are used as outputs to train the MLP regression model. To optimize the training effect, the data is cut near the determined value of the protection action probability value, and only a shorter length of data segment is retained for each fault location as the MLP regression model training set; The protection action state matrix after cutting is: Where: l is the length of the data segment reserved for each fault location.
6. The protection evaluation method for the protection action boundary value of the flexible DC system based on the multi-layer perceptron MLP according to claim 1 is characterized by: The method of calculating the transition resistance boundary value and the fault position boundary value of the protection action boundary based on the regression model of fault information and protection action is specifically as follows: when it is determined that the protection action probability value is at the set value and is at the protection action boundary, the fault position boundary value corresponding to a certain transition resistance under the protection action boundary is calculated using the regression model of fault information and protection action; and the transition resistance boundary value corresponding to a certain fault position under the protection action boundary is calculated.
7. A protection evaluation device for a flexible DC system protection action boundary value based on a multi-layer perceptron MLP, characterized in that: include: Data collection module, used to obtain the total training set and test set; The data preprocessing module is used to clean the total training set and test set data and clarify the protection action boundaries; MLP classification model module, which uses the MLP classification model to establish a prediction model for protection action behavior, and replaces discrete protection action behavior results with continuously changing protection action probability values to achieve a fitable result for protection action behavior; The MLP regression model module uses the MLP regression model to establish a regression model of fault information and protection action based on the results of the protection action behavior that can be fitted, and establishes a nonlinear functional relationship between fault information and protection action behavior, thereby solving the protection action boundary; the transition resistance boundary value and the fault position boundary value of the protection action boundary are calculated according to the regression model of fault information and protection action; The data output module evaluates the protection action performance according to the transition resistance boundary value and the fault position boundary value of the protection action boundary.
8. The protection evaluation device for the protection action boundary value of the flexible DC system based on the multi-layer perceptron MLP according to claim 7 is characterized in that: The MLP classification model module includes a regression model building module and a boundary value calculation module; The regression model building module uses the protection action probability value, fault location, and transition resistance value as input, and the transition resistance boundary value and fault location boundary value as output to train the MLP regression model. To optimize the training effect, the data is cut near the determined value of the protection action probability value, and only a shorter length of data segment is retained for each fault location as the MLP regression model training set; The protection action state matrix after cutting is: Where: l is the length of the data segment reserved for each fault location; The boundary value calculation module calculates the transition resistance boundary value and the fault position boundary value of the protection action boundary, determines that the protection is at the protection action boundary when the protection action probability value is the set value, and uses the regression model of fault information and protection action to calculate the fault position boundary value corresponding to a certain transition resistance under the protection action boundary; calculates the transition resistance boundary value corresponding to a certain fault position under the protection action boundary.
9. 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 protection evaluation method for the protection action boundary value of a flexible DC system based on a multi-layer perceptron MLP as described in any one of claims 1 to 6 are implemented.
