A Cable Fault Classification Method Based on Bagging-Heterogeneous k-Nearest Neighbor Boosting Learning

Abstract

The present invention discloses a method for classifying cross-bonded cable faults based on heterogeneous boosting learning. First, various faults such as joint faults, cross-bonding box faults, and abnormal grounding systems are simulated by PSCAD / EMTDC to obtain relevant electrical parameters such as main core and sheath currents, voltages, and grounding resistance values. Through parameter normalization, an electrical parameter feature matrix is constructed. Then, based on the k-nearest neighbor algorithm, different k values and different distance metrics are used as individual learners to construct a differential-height sub-learner. By introducing the Bagging algorithm, the overall learning efficiency of the heterogeneous learner is improved, and the distinction of cross-bonded transmission cable faults of different types and different regions can be realized. Compared with traditional classification methods such as SVM, k-nearest neighbor, and logistic regression, this method reduces the multi-classification error rate and spatial complexity and improves the generalization ability of the model, having great potential for engineering applications.

Description

Technical Field

[0001] The present invention relates to the technical field of power system transmission and distribution, and in particular to a cross-connected cable fault classification method based on Bagging-heterogeneous k-nearest neighbor boosting learning. Background Technique

[0002] With the continuous development of China's social economy and the continuous increase in power market demand, high-voltage, long-distance, and large-capacity power transmission technologies have also been steadily developed. Due to advantages such as small land occupation, little influence by environmental factors, and stable power transmission performance, power cables have gradually replaced overhead transmission lines and are widely used in power transmission and distribution lines. In particular, cable lines represented by cross-linked polyethylene (XLPE) insulated cables are widely used in power transmission and distribution systems.

[0003] When a power transmission and distribution cable fails, it is often accompanied by an increase in the induced voltage and induced current of the metal sheath, an increase in the temperature at the fault and joints, and the generation of partial discharge signals. Each type of fault has different fault characteristic signals. Therefore, by collecting relevant parameters such as sheath current, induced voltage, and grounding resistance in cross-connected boxes, protective grounding boxes, etc., and after certain processing, they can be used to distinguish cable fault types, and further judge the operating state of the cable, so as to achieve cable fault diagnosis.

[0004] Currently, in the application of cable fault classification, algorithms such as SVM, k-nearest neighbor, decision tree, and neural network are often used to classify and identify various types of data and pictures. However, these algorithms have a high dependence on raw data and low learning accuracy. In a simple SVM, the separation of the hyperplane is achieved by constructing the maximum margin. However, for large data, its processing performance is weak and it is only applicable to binary classification models, which is not conducive to multi-classification; the k-nearest neighbor algorithm can solve the multi-classification problem of large samples, but its generalization ability and approximation error are highly dependent on the k value, and the k value needs to be balanced well, otherwise its misclassification rate is relatively large. Other algorithms also have similar problems. Although the comprehensive performance of the improved classification algorithms has been improved to a certain extent, their application scenarios are relatively single. Summary of the Invention

[0005] Aiming at the current problems of incomplete classification and high misclassification error rate of cross-connected cable line faults in power transmission and distribution, this paper proposes a cross-connected cable fault classification method based on Bagging-heterogeneous k-nearest neighbor boosting learning. First, through PSCAD / EMTDC simulation of various faults such as joint faults, cross-connection box faults, and abnormal grounding systems, relevant electrical parameters such as main core and sheath currents, voltages, and grounding resistance values are obtained, and an electrical parameter feature matrix is constructed through parameter normalization. Then, based on the k-nearest neighbor algorithm, different k values and different distance metrics are used as individual learners to construct differential sub-learners, and the overall learning efficiency of heterogeneous learners is improved by introducing the Bagging algorithm. Compared with traditional classification methods such as SVM, k-nearest neighbor, and logistic regression, this method reduces the multi-classification error rate and space complexity and improves the generalization ability of the model. The proposed improved method has a faster speed and higher accuracy, and has great potential for engineering applications.

[0006] A cross-connected cable fault classification method based on Bagging-heterogeneous k-nearest neighbor boosting learning includes the following steps:

[0007] Step 1: Obtain the main core current, sheath current, and grounding resistance value of the cross-connected cable transmission line;

[0008] The above Step 1 can rely on PSCAD / EMTDC to simulate relevant parameters during normal and faulty conditions of the power transmission and distribution line;

[0009] Furthermore, use ammeters and voltmeters to obtain the main core currents and voltages at the beginning and end of the cable, the grounding currents at the beginning and end, and the sheath currents and voltages at the cross-connection points to obtain the initial parameter matrix D, and normalize the matrix to obtain D * ;

[0010] Furthermore, perform electrical processing on the above parameter features to obtain the feature parameter matrix X;

[0011] Furthermore, obtain the sample normalization parameter matrix T, and at this time, T is a special matrix containing various fault sample data.

[0012] Step 2: Construct fault labels

[0013] For the special matrix of fault sample data obtained in Step 1, it can respectively represent main core faults, sheath faults, grounding system faults, etc. of the cable, and its fault type is determined by the elements in the normalized parameter matrix;

[0014] Furthermore, classify the feature parameter matrix X corresponding to each type of fault, that is, the specific cable fault corresponds to the matrix X; at this time, the faults can be divided into N categories, corresponding to γ in the T matrix;

[0015] Furthermore, the Correspond one by one with specific simulation sample data.

[0016] Step 3: Construct a Bagging-Heterogeneous k-Nearest Neighbor Boosting Learning Model based on the sample matrix T; in the said Step 3, first construct a basic k-NN model, and through the sample matrix, construct a k-nearest neighbor model to be used as the later basic model and control group;

[0017] Furthermore, through the Bagging method, further process the original samples to obtain a new sample T'. Among them, when performing S times through the Bagging method, S + 1 training matrices with the same sample size as the original sample T matrix can be obtained;

[0018] Furthermore, taking a single sample matrix T as an example, train the heterogeneous k-NN model sequentially;

[0019] Furthermore, by evaluating the sample accuracy rate and recall rate, calculate the corresponding weights of each heterogeneous k-NN;

[0020] Furthermore, obtain the fault type classification and prediction function.

[0021] Beneficial effects: The present invention first constructs a parameter sample matrix, and then processes the sample data based on the Bagging-Heterogeneous k-Nearest Neighbor Boosting Learning method to realize the fault classification of 110 kV cross-connected lines. Greatly improves the problems of low accuracy and high complexity of traditional fault detection methods. At the same time, this method has strong generalization ability and can be applied to transmission lines of other levels and structures through appropriate modification. Brief Description of the Drawings

[0022] Figure 1 Cross-connected cable sheath connection diagram

[0023] Figure 2 Bagging-Heterogeneous k-Nearest Neighbor Boosting Learning Method

[0024] Figure 3 Bagging-Heterogeneous k-Nearest Neighbor Boosting Learning Fault Classification Flowchart

[0025] Figure 4 Simulation comparison diagrams of various faults Specific Embodiments

[0026] The following specifically expands the description of this method:

[0027] As Figure 1 shown, a cross-connected cable fault classification method based on Bagging-Heterogeneous k-Nearest Neighbor Boosting Learning includes the following steps:

[0028] Step 1: Obtain the main core current, sheath current and grounding resistance value of the cross-connected cable transmission line, denoted as matrix D = [UCX U PX I CX I PX R E ] T 5×5 (X=A, B, C), where U CX and I CX is the voltage and current of the X-phase main core at the beginning and end, U PX and I PX is the sheath induced voltage and sheath current, R E is the grounding resistance value (the line adopted is a single-ended grounding form at the beginning and end, and a cross-connected grounding in the middle. When the line structure changes, it is only necessary to change the structure of the corresponding matrix D. No special explanation will be given later).

[0029] The matrix D parameters in step 1 can be realized by using PSCAD / EMTDC to simulate the normal and faulty transmission and distribution lines, that is, respectively simulating main core faults (single core break and grounding), intermediate joint faults, sheath damage, interconnection box faults (wiring errors, water ingress, protector damage, etc.), and grounding system faults (multi-point grounding, floating grounding, etc.). By simulating these faults, the corresponding parameter matrix is obtained;

[0030] Furthermore, the matrix D is normalized, the main core current and voltage are converted to their per-unit values based on the rated current and voltage; the grounding resistance is converted to the per-unit value based on 4 ohms as the reference value (GB 50169-2006); the sheath induced voltage is converted to the per-unit value based on 50V as the reference voltage (GB 50127-2007), and the sheath current is converted to the corresponding per-unit value based on 12% of the rated current as the reference value. The above parameters are combined to reconstruct the normalized matrix, and we have

[0031] Furthermore, the parameter matrix is constructed from the normalized matrix At this time, the input feature matrix is T={(x1,γ1),(x2,γ2),...,(x N ,γ N )},in As input feature parameters, As the fault category, i=1,2…,N is the number of instance points.

[0032] Step 2: Determine the fault type through the parameter matrix

[0033] 1. When the feature matrix When the value changes, it is a main core fault. The main core fault can be divided into several categories. When the middle joint is broken down or the single-phase metallic grounding occurs, the value is infinite; but when the non-metallic grounding occurs, the value is a certain value; when the two lines are short-circuited, the value is another fixed value, etc. Corresponding to different main core faults, the fixed value is also different. Determine the fault type of the main core of the value judgment;

[0034] 2. When remains unchanged, changes, it is a ground system fault or an interconnection box fault at this time. When changes, it is a ground system fault at this time. Correspondingly, when remains unchanged, while the matrix has a single value change, it indicates that the sheath has a fault. When and both change, it is necessary to separately judge the change rates of the two.

[0035] 3. Through the above characteristic matrix, the fault characteristics of the 110 kV cross-connected line can be obtained. At this time, only the fixed value needs to be determined to determine the fault type. Further, weights can be assigned to various faults. For example, when it is the first type of main core fault, the weight value is 1, and the weight value is 0, and so on.

[0036] Step 3: Construct a Bagging-Heterogeneous k-Nearest Neighbor Boosting Learning Model based on the sample matrix T. The second step can only determine the basic fault type. Further, the fixed values of various faults can be determined through machine learning methods. First, a basic k-NN model needs to be constructed. In this model, the distance metric uses the traditional Euclidean distance, that is, The decision rule uses the majority voting method, that is, the class of the majority of the k nearest neighbors determines the class of the fault sample data point; the value of k is first taken as the general value 3, and then gradually improved until the optimal value k1 of the model is obtained, and a fault classification model with k = 3 is obtained as the later Bagging-Heterogeneous k-Nearest Neighbor control group.

[0037] Further, through the Bagging method, further processing is carried out to obtain a new sample T'. The sampling method of this sample is through the bootstrap method, that is, it is drawn from the sample matrix T containing N samples. Each time one is drawn, and it is drawn N times, and then the new sample matrix T' can be obtained, T' = {(x'1,γ'1),(x'2,γ'2),...,(x' N ,γ' N )}. Through the Bagging method, more new sample matrices can be obtained. In this paper, S = 3 is taken, that is, there are a total of 4 sample matrices. Due to the particularity of the Bagging method, at this time, about 36.8% of the samples are not in the newly constructed samples. At this time, these samples can be used as the "out-of-bag estimate" of the generalization ability of the new model;

[0038] Furthermore, a heterogeneous k-NN model is constructed, which is a k-NN ensemble learner under different k values and different types of distance metrics. This k-NN ensemble learner constructs a "strong learner" from "weak learners" based on Bagging-heterogeneous k-nearest neighbors. At this time, distance metrics such as "Manhattan distance" or "Mahalanobis distance" can be used. In this paper, two groups of Euclidean distances and two groups of Mahalanobis distances are used as distance metrics.

[0039] Furthermore, the optimal k value is obtained. At this time, the k value is first taken as the general value 3. After one round of training, the classification error rate F is obtained. When the sample is a fault sample and the model is classified correctly, it is recorded as TP; when classified incorrectly, it is recorded as FP; when the sample is a non-fault sample and the model is classified correctly, it is recorded as FN, and when classified incorrectly, it is recorded as TN. We can get:

[0040]

[0041] At this time, to improve the fault classification accuracy rate, attention should be paid to misclassified samples, and the k value is modified again. Denote the correct rate of each type of fault as η, then its positive rate is η = [η1 η2... η K , and the maximum diagnostic accuracy rate, that is, η max , is obtained, and the corresponding k value at this time is the optimal k value.

[0042]

[0043] The optimal k of this k-NN is obtained separately O = [k O1 k O2 , and for the Mahalanobis distance group, the optimal value k M = [k M1 k M2 needs to be obtained separately.

[0044] Furthermore, the classification characteristics of the Mahalanobis distance group and the Euclidean distance group in different faults are obtained. For the k-NN groups where the Mahalanobis distance group is classified correctly and the Euclidean distance group is classified incorrectly, the weight of the Mahalanobis group model needs to be increased. Furthermore, for the trained heterogeneous k-NN, the fault judgment model is obtained:

[0045]

[0046] In the formula is the fault value predicted by each classifier, and ω i is the weight of each classifier. At this time, as long as the corresponding electrical value x i is input, it can be determined whether the data is fault data, and the fault type can be determined by the value of H(X). This method is also applicable to other voltage level transmission lines, and only the normalization matrix reference value needs to be changed.

[0047] Finally, it should be noted that the above method only takes the failure of an 110 kV cross-connected line as an example to illustrate a cross-connected cable fault classification method based on Bagging-heterogeneous k-nearest neighbor boosting learning. However, this method is not limited to this case. Any method involving similar extraction of electrical characteristic parameters, normalization, and data processing should be regarded as the spirit and principle of the present invention. Partial modifications, substitutions, and improvements made to this method should be included in the claims of the present invention and should be protected by the present invention.

Claims

1. A cross - interconnected cable fault classification method based on Bagging - heterogeneous k - nearest neighbor boosting learning, characterized in that By detecting the voltages and currents at the beginning and end of the main cores of 110 kV cross-connected lines or other high-voltage cable transmission lines, as well as the voltages, currents of the sheath in the cross-connected sections, and the grounding resistance values of each section, the operating conditions of the cables can be determined; By constructing the initial parameter matrix D, at this time, the rated voltage, rated current, etc. of the corresponding transmission line are used as the reference values and normalized to obtain D * , and then construct the characteristic parameter matrix X: where are the per-unit values of the main core voltage and current at the head and tail ends, are the per-unit values of the sheath voltage and current in the cross-bonding section, is the per-unit value of the grounding resistance; When there is a change in the characteristic matrix it indicates a main core failure. The main core failure can be divided into several types. When the intermediate joint breaks down or there is a single-phase metallic ground, this value is infinite. However, when there is a non-metallic ground, this value is a certain value; when there is a short circuit between two wires, it is another certain value, etc. Corresponding to different main core failures, the fixed values are different. Therefore, the type of main core failure can be judged by the value When remains unchanged changes, it indicates a ground system fault or an interconnection box fault. When changes, it indicates a ground system fault. Correspondingly, when remains unchanged, while the matrix has a single - value change, it indicates a sheath fault. When and both change, it is necessary to separately judge the change rates of the two; Further, a parameter matrix is constructed from the normalization matrix At this time, the input feature matrix has T = {(x1, γ1), (x2, γ2),..., (x N , γ N )}, where As input feature parameters, As the fault category, i = 1, 2…, N is the number of instance points; At this time, the sample normalization parameter matrix T is used as the training set and the test set. The Euclidean distance is selected as the distance metric, and the initial k value is set to the general value 3. Start training the kNN model, and through the Bagging method, process the original samples to obtain new samples T'. Further construct a heterogeneous k-NN model, and this heterogeneous k-NN model is the k-NN ensemble learner under different k values and different types of distance metrics; at this time, the k value and the distance metric in the kNN model change, calculate the fault classification accuracy rate of this heterogeneous k-nearest neighbor boosting learning method, and find the maximum accuracy rate η max , and calculate the corresponding k value at this time, which is the optimal k value.

2. For a cross-connected cable fault classification method based on Bagging-heterogeneous k-nearest neighbor boosting learning according to claim 1, when a fault occurs in the cable transmission line, the feature matrix changes to a certain extent. At this time, as long as the fault feature values of the target line are determined, the fault type can be determined.

3. For a cross-connected cable fault classification method based on Bagging-heterogeneous k-nearest neighbor boosting learning according to claim 1, at this time, the fault feature values can be obtained based on Bagging-heterogeneous k-nearest neighbor boosting learning, that is, by constructing a sample normalization parameter matrix T and marking the fault labels for various samples. At this time, the faults can be divided into N categories, corresponding to γ in the T matrix.

4. For a cross-connected cable fault classification method based on Bagging-heterogeneous k-nearest neighbor boosting learning according to claim 1, calculate the fault weights in the Bagging-heterogeneous k-nearest neighbor boosting learning method to obtain a fault judgment model: where is the fault value predicted by each type of classifier, and ω i is the weight of each type of classifier; at this time, as long as the corresponding electrical value is input, it can be determined whether the data is faulty data, and the fault type can be determined by the value of H(X).

Images