A power distribution automation terminal fault positioning method based on a double dimension reduction algorithm

Abstract

The application discloses a power distribution automation terminal fault positioning method based on a double dimension reduction algorithm, and the method steps are as follows: firstly, analog quantity, switch action behavior information, configuration information and device self-checking alarm information are read from a power distribution automation system as characteristic samples for fault identification; secondly, random forest RF is used to filter the characteristic samples, and an input matrix of principal component analysis PCA is optimized; then, the input matrix is reduced in dimension by PCA, and the complexity of the input matrix is reduced; meanwhile, the number of principal components of PCA, a kernel parameter and a penalty factor of a support vector machine SVM are optimized by using a particle swarm; finally, an RF-PCA-SVM classification model is constructed by using the optimized parameters, a fault label is obtained, and a fault point is found out by comparing a fault positioning matrix. The application can accurately and efficiently process high-dimensional data samples, and greatly reduces the calculation time consumption on the premise of ensuring the accuracy.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of fault maintenance of power distribution automation terminals, and particularly relates to a power distribution automation terminal fault positioning method based on a two-dimensional dimension reduction algorithm BACKGROUND

[0002] A power distribution network contains a large number of power distribution automation terminals, and up to ten thousand terminals can be simultaneously operated in a power distribution network of a large city. Ensuring that the power distribution automation terminals are in a good working state is an important basis for good operation of the power distribution automation system. Most of the power distribution automation terminals are directly operated outdoors or in a simple sheltered environment, and compared with substation automation devices, the working environment is poor and is more susceptible to environmental factors. Therefore, the power distribution automation terminals have more stringent requirements on their adaptive temperature and humidity ranges, moisture resistance, shock resistance, and magnetic resistance. However, the power distribution automation terminals have various types and manufacturers, and the product quality is uneven, and various maintenance tools result in very low operation reliability, and abnormal problems often occur without a clear cause. In addition, the power distribution automation terminals are generally installed at the periphery of power distribution primary equipment, and are distributed in a wide range, which makes the daily operation and maintenance work very difficult.

[0003] In recent years, the State Grid Corporation of China has continuously promoted the construction of power distribution automation, but as the construction scale of the power distribution automation rapidly expands, the related operation and maintenance technologies and means are still relatively backward. At present, the operation and maintenance of the power distribution network still adopts two modes of fault maintenance and periodic maintenance. The fault maintenance mode generally adopts simple replacement maintenance, the equipment with faults is not repaired but directly replaced, and the operation and maintenance cost is extremely high. The periodic maintenance mode adopts manual on-site inspection, and a large amount of labor is consumed and the efficiency is low. The power distribution automation terminals are large in number and complex in operation environment, and further increase the difficulty of operation and maintenance. To realize rapid and accurate positioning of the faults of the power distribution automation terminal equipment, the research on intelligent fault positioning technology is the key to solving the problem. However, the power distribution automation monitoring data are continuously expanded in scale and increased in quantity, and the single intelligent fault positioning method performs poorly. SUMMARY

[0004] The application aims to provide a power distribution automation terminal fault positioning method capable of accurately and efficiently processing high-dimensional data samples, and greatly improving the calculation speed on the premise of ensuring the fault positioning accuracy.

[0005] The technical solution for realizing the application is as follows: a power distribution automation terminal fault positioning method based on a two-dimensional dimension reduction algorithm, comprising the following steps:

[0006] Step 1, from the power distribution automation system, collect the historical faults of the power distribution automation terminal, generate a fault positioning matrix according to the fault type according to the position of the fault point, and generate data samples for fault positioning from the analog quantity information of the fault related equipment, the switch action behavior information, the configuration information and the device self-checking alarm information, and these data samples form a fault database;

[0007] Step 2, dimension reduction processing is performed on the fault database:

[0008] Step 21, a decision tree is constructed, the importance of each dimension feature of the fault data is evaluated through the error rate of out-of-bag data OOB, and the optimal feature subset is screened out;

[0009] Step 22, the feature value decomposition is performed on the feature subset, the transformation matrix is obtained according to the size of the feature value, and the PCA matrix containing all principal components is generated;

[0010] Step 3, the number of principal components k of the PCA matrix, the penalty factor C of the support vector machine and the kernel parameter g are taken as individuals in the particle swarm algorithm population, the sum of the correct rate of the PCA matrix classification result and the calculation time of the SVM classifier is taken as the fitness function, and a fault positioning model is constructed;

[0011] Step 4, the optimal solution of the fault positioning model is the optimal parameter, the fault positioning model is trained according to the optimal parameter, and the RF-PCA-SVM fault positioning model is obtained;

[0012] Step 5, when the power distribution terminal fails, the analog quantity information of the fault related equipment, the switch action behavior information, the configuration information and the device self-checking alarm information are taken as inputs, the current fault label is obtained, and the fault point is found out by referring to the fault positioning matrix of step 1.

[0013] Compared with the prior art, the present application has the following advantages: (1) the analog quantity, switch action behavior information, configuration information and device self-checking alarm information are read from the power distribution automation system as the feature samples for fault recognition, the feature samples are screened by using the random forest RF, and the input matrix of the principal component analysis PCA is optimized; (2) the input matrix is reduced in dimension by using the PCA, the complexity of the input matrix is reduced, and the principal component number of the PCA, the kernel parameter and the penalty factor of the support vector machine SVM are optimized; (3) the RF-PCA-SVM classification model is constructed by using the optimized parameters, the fault label is obtained, and the fault point is found out by referring to the fault positioning matrix, so that the high-dimensional data samples can be accurately and efficiently processed, and the calculation time is greatly reduced under the premise of ensuring the accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0014] Figure 1A flow chart of the power distribution automation terminal fault positioning method based on the double dimension reduction algorithm.

[0015] Figure 2 A flow chart of the particle swarm optimization algorithm.

[0016] Figure 3 A random forest feature subset identification accuracy curve.

[0017] Figure 4 A particle swarm optimization algorithm iteration convergence curve. DETAILED DESCRIPTION

[0018] As Figure 1 shown, the power distribution automation terminal fault positioning method based on the double dimension reduction algorithm includes the following steps:

[0019] Step 1, from the power distribution automation system, collect the power distribution automation terminal historical fault, generate a fault positioning matrix according to the fault type according to the fault point position, generate a data sample for fault positioning from the analog quantity information, switch action behavior information, configuration information and device self-check alarm information of the fault related equipment, and these data samples form a fault database;

[0020] Step 2, dimension reduction processing is performed on the fault database:

[0021] Step 21, a decision tree is constructed, the importance of each dimension feature of the fault data is evaluated through the out-of-bag data OOB error rate, and the optimal feature subset is selected;

[0022] Step 22, the feature value decomposition is performed on the feature subset, the transformation matrix is obtained according to the size of the feature value, and the PCA matrix containing all principal components is generated;

[0023] Step 3, the number of principal components k of the PCA matrix, the penalty factor C of the support vector machine and the kernel parameter g of the three variables are taken as individuals in the particle swarm algorithm population, the sum of the correctness rate of the PCA matrix classification result and the calculation time of the SVM classifier is taken as the fitness function, and a fault positioning model is constructed;

[0024] Step 4, the optimal solution of the fault positioning model is the optimal parameter, the fault positioning model is trained according to the optimal parameter, and the RF-PCA-SVM fault positioning model is obtained;

[0025] Step 5, when the power distribution terminal fails, the analog quantity information, switch action behavior information, configuration information and device self-check alarm information of the fault related equipment are taken as inputs, the current fault label is obtained, and the fault point is found by referring to the fault positioning matrix of step 1.

[0026] As a specific example, step 1 generates the original data sample, as follows:

[0027] From the power distribution automation system, the analog quantity information, switch action information, configuration information and device self-check alarm information of the relevant equipment in the historical fault of the power distribution automation terminal are collected to generate a data sample for fault location. A fault location matrix is generated according to the fault occurrence point. Taking a station terminal (DTU) as an example, the design rules of the fault location matrix are as follows:

[0028] D4-D0: Determine which position of the device the fault occurrence point is in. For example, the remote signaling or remote current fault is 10000, the remote voltage fault is 01000, the power supply fault is 00110, and the communication fault is 00001.

[0029] D8-D5: Determine whether the fault is caused by short circuit, open circuit, reverse connection or point table error. Fill in 1 for yes and 0 for no. If both reasons are possible, fill in 1 in both columns.

[0030] D10-D9: Since there are remote signaling input loops and remote control output loops in the 26-core navigation plug, D10 and D9 are used to distinguish the same type of faults occurring in the 26-core navigation plug. The output loop is 01 and the input loop is 10.

[0031] D14-D11: Determine whether the fault point is a voltage terminal, a current terminal, a zero sequence terminal or an N terminal. Fill in 1 for yes and 0 for no.

[0032] D15: Determine whether the device is in a running state. Fill in 1 for yes and 0 for no.

[0033] The fault location matrix is shown in Table 1:

[0034] Table 1 Fault location matrix

[0035]

[0036] In order to facilitate the operation and maintenance personnel to quickly locate the fault occurrence point, the fault type and the corresponding fault point position are connected through the fault location matrix. The fault point and fault label correspondence is shown in Table 2. Binary numbers and hexadecimal numbers are combined to encode, which can quickly locate the fault when processing the fault. Four hexadecimal numbers can represent hundreds of fault types, which meets the needs of the number of possible fault types.

[0037] Table 2 Fault location table

[0038]

[0039] As a specific example, step 2 performs dimensionality reduction on the original data samples, specifically as follows:

[0040] Step 21: Train each decision tree by random sampling, calculate the contribution value of each dimension feature on each tree, then take the average value as the importance of the feature, and finally filter out a feature subset according to the importance of each dimension feature.

[0041] The function Imp for evaluating feature importance i , as shown in equation (1):

[0042]

[0043] Where N represents the number of decision trees constructed; errOOB1 j represents the out-of-bag data error of the jth decision tree; errOOB2 j,i represents the out-of-bag data error of the jth decision tree after adding a certain disturbance to the ith feature (modifying the value of the sample at feature i);

[0044] Step 22: Perform eigenvalue decomposition on the feature subset to obtain a transformation matrix according to the eigenvalue size, and generate a PCA matrix containing all principal components;

[0045] The solution formula of the PCA matrix is represented as:

[0046] M' = MV (2)

[0047] Where V represents the corresponding first k eigenvectors (V1, V2, …, V k ) after sorting by eigenvalue; M represents the optimal feature subset obtained in step 21; M' represents the PCA matrix generated from M;

[0048] As a specific example, step 3 finds the optimal parameters for building the model, specifically as follows:

[0049] Optimize the RF-PCA-SVM classifier using the PSO algorithm, take the number of principal components of the PCA matrix, the penalty factor C of the SVM, and the kernel parameter g as the individuals in the particle swarm algorithm population, and take the sum of the accuracy and the calculation time of the SVM classification result on the PCA matrix as the fitness function. The optimal solution is the optimal parameter for building the fault location model. The particle swarm optimization steps are as Figure 2 shown, including the following five sub-steps:

[0050] Step 31: Initialize parameters, randomly generate the coordinates and velocities of particles

[0051] Set SVM kernel function, and set the value range of SVM penalty factor C, kernel function parameter g and principal component number k respectively, set X i = (x i1 , x i2 , x i3 ) as the coordinate solution of the optimal model parameters, and X i is called a particle, X ij indicates the coordinate solution of the jth parameter in the ith particle, and a population X = (X1, X2, …, X m ) is composed of m particles; in the iterative calculation, the speed of each particle is recorded as V i = (v i1 , v i2 , v i3 ), wherein v ij indicates the speed of the jth parameter coordinate of the ith particle.

[0052] Step 32: training SVM classifier and calculating fitness

[0053] The hyperplane is defined as formula (3).

[0054]

[0055] In the formula, indicates the kernel function; C represents the penalty factor, the greater the value of C, the fewer the error samples allowed; ζ i is a slack variable, indicating an error sample variable; ω is the normal vector of the hyperplane; b is the bias; y i indicates the actual target value, y i ∈{-1,1}.

[0056] The optimal hyperplane problem is combined with the dual problem, and the classification function f(x) is obtained as formula (4).

[0057]

[0058] In the formula, indicates the Lagrange multiplier; according to f(x) = 1 or f(x) =-1, two types of faults are distinguished;

[0059] By designing an SVM classifier between any two types of samples, multiple fault classifications are realized, if the samples need to be divided into k categories, then k(k-1) / 2 classifiers are designed, and the final result is obtained by voting of all classifiers.

[0060] The fitness solving formula can be expressed as:

[0061]

[0062] In the formula: is the label of the SVM model classification result; y i is the actual label of the sample; k is the number of principal components selected (i.e., the dimensionality after dimensionality reduction); n is the number of fault samples;

[0063] Step 33: updating the extreme value of each particle and the extreme value of the population according to the fitness value;

[0064] Step 34: updating the speed and coordinates of the current particle: the particle will update the speed and coordinates according to the coordinates, speed, individual optimal position and population optimal position of the last time in each iteration, and the updating formula is shown in formula (6) and formula (7).

[0065]

[0066] In the formula, u is the inertia weight; t is the current iteration number; is the particle position of the t-1th iteration; is the particle speed of the tth and t-1th iteration, respectively; is the individual optimal position of the t-1th iteration; gbest t-1 is the population optimal position of the t-1th iteration; l1 represents the individual learning factor, and the greater the individual factor, the more the particle tends to the optimal value of the particle; l2 represents the social learning factor, and the greater the social learning factor, the more the particle tends to the optimal value of the population; r1 and r2 are random numbers in the interval [0, 1], which are used to increase the randomness of the search;

[0067] Step 35: when the iteration number is reached or the fitness condition is met, the iteration is terminated; otherwise, return to step 32.

[0068] As a specific example, the model is trained with the optimal parameters obtained in step 4 to perform fault positioning, which is specifically as follows:

[0069] The fault positioning model is trained according to the optimal parameters, and the fault positioning model is obtained. When a fault occurs in the power distribution terminal, the related analog quantity information, switch action behavior information, configuration information and device self-check alarm information are used as inputs to determine the current fault type, and the fault point is found by referring to the fault positioning matrix of step 1.

[0070] The application will be further described in detail in combination with the embodiments.

[0071] Embodiment

[0072] In order to verify the effectiveness of the scheme of the application, the following simulation experiment is performed.

[0073] Step 1: generating original data samples

[0074] The fault simulation experiment is designed on the distribution automation experiment platform of Shanghai Jinzhi Power Technology Co., Ltd. The data of 20 common fault types of the station terminal (DTU) are collected. The characteristic information is preliminarily screened, and finally 3120 groups of 126 fault data characteristics are obtained.

[0075] Step 2: Dimensionality reduction processing is performed on the original data sample

[0076] Step 2.1: The number of decision trees N of the random forest is set to 50. The original data is randomly sampled 126 times with replacement, 50 groups of training data are generated, 10 times of feature importance calculation are performed according to formula (1) and the average value is taken, the importance is sorted according to the importance, and part of the feature importance is shown in Table 3.

[0077] Table 3 Part of feature importance

[0078]

[0079] In order to compare the influence of each feature on the fault recognition accuracy, one feature with the smallest importance is removed in turn according to the feature importance, the standard SVM model is taken as the reference, the fault classification accuracy under the corresponding feature subset is calculated, and finally the feature subset with the highest accuracy is selected from the feature set.

[0080] The recognition accuracy curve of the feature subset is shown in Figure 3 It can be seen from Figure 3 that when all the 126-dimensional features are selected as input, due to the existence of many irrelevant variables and redundant features, the recognition accuracy is only 83.45%, when the first 46-dimensional features are selected as input, the fault classification accuracy is the highest, reaching 91.03%, and the optimal feature subset is obtained.

[0081] Step 2.2: PCA is used to perform feature decomposition on the optimal 46-dimensional feature subset, the feature vectors are arranged according to the feature value size to obtain the projection matrix V, and the sample matrix M' after dimensionality reduction is obtained according to formula (2).

[0082] Step 3: Finding the optimal parameters for model construction

[0083] The population number of PSO is set to 10, the maximum iteration number is set to 100, l1 and l2 are set to 2, the optimization range of the penalty parameter C is [0.001, 100], the optimization range of the kernel parameter g is [0.001, 100], the optimization range of the principal component number k is [1, 46], and the iterative optimization is performed according to formula (6) and formula (7). The convergence result is shown in Figure 4 The optimal parameters {C, g, k} obtained finally are {166.11, 0.7432, 23}.

[0084] Step 4: training the model with the obtained optimal parameters to perform fault location.

[0085] The optimal RF-PCA-SVM model is constructed by using the optimal principal component number k and the optimal parameters C and g, 70% of the reduced samples are used to train the model, and the remaining 30% of the data are used to verify the algorithm performance. The SVM, RF-SVM, PCA-SVM and the method of the application are used for fault classification, and the classification results are shown in Table 4. It can be seen from the table that the dimension reduction model based on RF-PCA can not only improve the accuracy of the fault location algorithm, but also reduce the input dimension and thus reduce the training time.

[0086] Table 4 Classification results

[0087]

[0088] In summary, the application reads analog quantity, switch action behavior information, configuration information and device self-checking alarm information from the power distribution automation system as the feature samples for fault identification, uses the random forest RF to screen the feature samples, optimizes the input matrix of the principal component analysis PCA, performs dimension reduction on the input matrix by using the PCA, reduces the complexity of the input matrix, simultaneously optimizes the principal component number of the PCA, the kernel parameter and the penalty factor of the support vector machine SVM by using the particle swarm, constructs the RF-PCA-SVM classification model by using the optimized parameters, obtains the fault label, finds out the fault point by checking the fault location matrix, and can accurately and efficiently process high-dimensional data samples, greatly reduces the calculation time on the premise of ensuring the accuracy.

Claims

1. A power distribution automation terminal fault location method based on a two-fold dimension reduction algorithm, characterized in that, Comprising the following steps: Step 1, from the power distribution automation system, collecting the power distribution automation terminal historical fault, generating a fault positioning matrix according to the fault type according to the fault point position, analog quantity information of fault related equipment, switch action behavior information, configuration information and device self-check alarm information, generating data samples for fault positioning, these data samples form a fault database; Step 2, dimension reduction processing is performed on the fault database: Step 21, a decision tree is constructed, the importance of each dimension feature of the fault data is evaluated through the error rate of out-of-bag data OOB, and the optimal feature subset is selected; Step 22, the feature subset is subjected to eigenvalue decomposition, and a transformation matrix is obtained by sorting according to the eigenvalue size, and a PCA matrix containing all principal components is generated; Step 3, taking the number of principal components of the PCA matrix k, the penalty factor C of the support vector machine and the kernel parameter g as individuals in the particle swarm algorithm population, taking the sum of the correct rate of the SVM classifier on the PCA matrix classification result and the calculation time consumption as the fitness function, and constructing a fault positioning model; Step 4, the optimal solution of the fault positioning model is the optimal parameter, and the RF-PCA-SVM fault positioning model is obtained by training the fault positioning model according to the optimal parameter; Step 5, when the power distribution terminal fails, the analog quantity information of the fault related equipment, the switch action behavior information, the configuration information and the device self-check alarm information are taken as inputs to obtain the current fault label, and the fault point is found out by referring to the fault positioning matrix of step 1.

2. The method of claim 1, wherein the method is based on a two-stage dimension reduction algorithm. In step 21, the function Imp that evaluates the importance of a feature is estimated for each decision tree by random sampling i As follows: where N denotes the number of constructed decision trees; errOOB1 j denotes the out-of-bag data error of the jth decision tree; errOOB2 j,i denotes the out-of-bag data error of the jth decision tree after adding a perturbation to the ith feature of the out-of-bag data, i.e., modifying the value of the sample at feature i.

3. The method of claim 1, wherein the method is based on a two-stage dimension reduction algorithm. In step 2.2, the eigenvectors are sorted according to the eigenvalue size, and the solution formula of the PCA matrix is represented as: M' = MV Where V represents the first k eigenvectors (V1, V2, ..., V) after sorting by eigenvalues. k M represents the optimal feature subset; M′ represents the PCA matrix generated by M.

4. The method of claim 1, wherein the method is based on a two-stage dimension reduction algorithm. In step 3, the classification method of the SVM classifier is as follows: Define the hyperplane as shown in the following formula: wherein represents a kernel function; C represents a penalty factor, the greater the value of C, the fewer the error samples allowed; ζ i is a slack variable, representing an error sample variable; ω is a normal vector of a hyperplane; b is a bias; y i represents an actual target value, y i ∈{-1,1}. Combine the optimal hyperplane problem with the dual problem to obtain the classification function f(x) as shown in the following formula: wherein denotes the Lagrange multiplier; two classes of faults are distinguished according to f(x) = 1 or f(x) = -1; By designing an SVM classifier between any two types of samples, multiple fault classifications are realized, if the samples need to be divided into k categories, then k(k-1) / 2 classifiers are designed, and the final result is obtained by voting of all classifiers.

5. The method of claim 1, wherein the method is based on a two-stage dimension reduction algorithm. In step 3, the specific method of optimizing parameters by particle swarm algorithm is as follows: Step 31, initialize parameters, randomly generate coordinates and speed of particles; Set SVM kernel function, and set the value range of SVM penalty factor C, kernel function parameter g and principal component number k respectively, set X i = (x i1 ,x i2 ,x i3 ) as the coordinate solution of the best model parameters, and call X i a particle, X ij indicates the coordinate solution of the jth parameter in the ith particle, and a population X = (X1, X2, …, X m ) is composed of m particles; in the iterative calculation, the speed of each particle is recorded as V i = (v i1 ,v i2 ,v i3 ), wherein v ij indicates the speed of the ith particle at the jth parameter coordinate; Step 32, train the SVM classifier and calculate the fitness, and the solution formula of fitness is represented as: In the formula, is the label of the model classification result; y i is the actual label of the sample; k is the number of principal components, i.e., the dimension after dimension reduction; n is the number of fault samples; Step 33, update the extreme value of each particle and the extreme value of the population according to the fitness value; Step 34, update the speed and coordinates of the current particle: each iteration particle updates the speed and coordinates according to the last coordinates, speed, individual optimal position and group optimal position, and the update formula is as follows: where u is the inertia weight; t is the current iteration number; is the particle position of the (t-1)th iteration; is the particle velocity of the tth and (t-1)th iteration, respectively; is the individual optimal position of the (t-1)th iteration; gbest t-1 is the population optimal position of the (t-1)th iteration; li represents the individual learning factor, the larger the individual factor, the more the particle tends to the optimal value of the particle; l2represents the social learning factor, the larger the social learning factor, the more the particle tends to the optimal value of the population; r1and r2are random numbers in the interval [0, 1] for increasing the randomness of the search; Step 35, when the iteration number is reached or the fitness condition is met, the iteration is terminated; otherwise, return to step 32.

6. The method of claim 1, wherein the method is based on a two-stage dimension reduction algorithm. In step 1, for the station terminal DTU, the fault positioning matrix is as shown in table 1, which is as follows: D4~D0: judge which position of the device the fault occurs, the remote signaling or remote current fault is 10000, the remote voltage fault is 01000, the power supply fault is 00110, and the communication fault is 00001; D8~D5: Determine whether the fault is caused by short circuit, open circuit, reverse connection or point table error; fill in 1 for yes and 0 for no, or fill in 1 for both if both reasons are possible; D10~D9: Since there are remote input circuits and remote output circuits in the 26-core navigation plug, D10 and D9 are used to distinguish the same type of faults occurring in the 26-core navigation plug, with output circuit being 01 and input circuit being 10; D14~D11: Determine whether the fault point is the voltage terminal, current terminal, zero sequence terminal or N terminal, fill in 1 for yes and 0 for no; D15: Determine whether the device is in running state, fill in 1 for yes and 0 for no; Table 1 Fault location matrix The fault type and the corresponding fault point position are connected through the fault location matrix, and binary numbers and hexadecimal numbers are combined for coding. When handling faults, positioning is performed, and the correspondence between fault points and fault labels is shown in Table 2: Table 2 Fault location table

Images