Intelligent line fault detection method for active power distribution network

Abstract

The invention discloses an intelligent line fault detection method for an active power distribution network, and belongs to the technical field of line fault detection, and the method comprises the steps: S1, taking the fault voltage and the voltage fundamental wave of each measurement point of an active power distribution network line as characteristic indexes for judging the operation state of the line; s2, extracting frequency band voltage peak values, inter-phase voltage differences and sixth harmonic components of line measurement points in normal and fault states as fault feature samples; s3, performing data preprocessing; s4, splitting the training set and the test set; s5, inputting the classified data set into a LightGBM (LightGBM) model for classification; s6, combining a plurality of independent decision tree classifiers in the LightGBM to generate a strong classifier; s7, inputting the output value of the strong classifier into the DNN for learning and training; s8, performing fault detection performance evaluation; and S9, outputting a fault detection result. The problems that after a large number of distributed power supplies are connected to the power distribution network, the system topological structure is complex, and line faults are difficult to locate are solved.

Description

Technical Field

[0001] The invention belongs to the technical field of line fault detection, and in particular relates to an intelligent line fault detection method for an active power distribution network. Background Art

[0002] With the large-scale access of distributed power sources to the power grid, the traditional distribution network has gradually evolved into an active distribution network structure with multiple distributed power supply endpoints. This change has significantly increased the transmission paths between the power supply and the load, the system topology has become more complex, and the line flow has become multi-directional, which has greatly increased the risk of line current circulation and imbalance, and the line is more prone to short-circuit faults. When a short-circuit fault occurs, the continued power supply of other distributed power sources in the line may cause the system to produce an "island effect" and continue to transmit power to the fault area, thereby threatening the safety of maintenance personnel and the normal operation of equipment. Therefore, how to accurately locate the fault location and identify the fault type has become a research hotspot in the current electrical engineering field.

[0003] Traditional distribution network line disconnect switches are usually reset every few weeks or even months. However, due to the large-scale access of intermittent distributed power sources, switches may need to be reconfigured frequently on an hourly basis. In this case, directional overcurrent relays are difficult to set in real time and are no longer suitable for the complex line structure of current active distribution networks. Therefore, the research on fault detection of active distribution networks containing multiple distributed power sources has gradually become a research focus in the field of distribution network fault detection.

[0004] At present, artificial intelligence technology is attracting extensive attention from researchers at home and abroad due to its advantages in data analysis and autonomous learning. Since artificial intelligence technology mainly relies on the operating data itself, when the distribution network structure changes, the impact on the classification, prediction, and identification functions of the model is relatively small. Therefore, it has been gradually applied in the field of distribution network fault detection. With the powerful learning and adaptive capabilities of artificial intelligence, the fault detection model can more flexibly respond to complex changes in the system structure, thereby improving detection accuracy and real-time performance. A method for intelligent line fault detection in active distribution networks is needed. Summary of the invention

[0005] The purpose of the present invention is to improve the accuracy of active distribution network line fault detection and the visualization of the fault diagnosis process, and to propose an intelligent fault detection method that combines artificial intelligence technology with distribution network line fault detection. The combined model uses the fault voltage and voltage fundamental wave of the measurement point as characteristic indicators, and extracts the voltage peak, phase-to-phase voltage difference and sixth harmonic component of each frequency band as fault feature samples. In this way, the model can accurately identify fault features, further improve the sensitivity and accuracy of fault detection, and provide visualization support for the fault diagnosis process.

[0006] The object of the present invention is achieved by: a method for detecting faults in an intelligent line of an active power distribution network, comprising the following steps:

[0007] S1. The fault voltage and voltage fundamental wave of each measuring point of the active distribution network line are used as characteristic indicators to judge the operation status of the line; S2. The frequency band voltage peak, phase-to-phase voltage difference and 6th harmonic component of the normal and faulty state of the line measurement point are extracted as fault feature samples; S3. Data preprocessing is performed, including denoising pre-operation of the data in S2; S4. The training set and test set are split, and the data set is divided into training set and test set in a ratio of 7:3; S5. The divided data set is input into the LightGBM model for classification; S6. Multiple separate decision tree classifiers in LightGBM are combined to generate a strong classifier; S7. The output value of the strong classifier is input into the DNN for learning and training; S8. Fault detection performance evaluation is performed; S9. The fault detection result is output. The classification results of the fault feature samples are input into the neural network as feature data for training and learning. The Shapley value analysis method is introduced in the model to determine the contribution of each fault feature to the prediction of the combined model when faults occur in different branches of the active distribution network, so as to realize the visualization of the influencing factors in the fault diagnosis decision. This method not only helps to reduce the amount of feature data collection required for the model, but also improves the operating efficiency of the model. Therefore, the present invention uses artificial intelligence technology based on the characteristic relationship between voltage and line fault to realize fault detection of active distribution network lines, thereby effectively promoting the intelligent and efficient operation of active distribution networks.

[0008] Furthermore, the LightGBM model includes extending the loss function through the second-order Taylor series, adopting the Leaf-Wise decision tree growth strategy, deeply limiting the data expansion process, shortening the training time, and reducing memory consumption, so as to reduce the risk of overfitting and improve the model operation efficiency; the implementation steps of the strong classifier generation part of LightGBM are as follows,

[0009]

[0010] In the formula, r ti It is the residual calculation value, which is used to measure the current model prediction value f t-1 (x i ) and y i Deviation of x i is the input characteristic value of the data sample of the i-th active distribution network line operation measurement point; y i is the true value of the operating measurement point of the i-th active distribution network line;

[0011] Fitting is performed based on the Leaf-Wise decision tree growth strategy, using (x i ,rti ) fits a CART and obtains the tth regression tree, whose corresponding leaf node region is R tm (m=1,2,…,M), where M is the number of leaf nodes, and for each leaf node, find the solution to minimize the loss function, that is, to fit the best output value h of the leaf node t (x),

[0012]

[0013] In the formula, h t (x) is the output value of the fitted leaf node; H is the hypothesis space of the weak learner, usually all possible CART decision trees; R tm is the region of the mth leaf node in the decision tree; h is the weak learner fitted in the current round;

[0014] Let I(x∈R tm ) is the input variable x in R tm The fitting function h of the decision tree is a simple parameterized function of the leaf node area. t (x) is,

[0015]

[0016] In the formula, c tm is the output value of the mth leaf node in the decision tree.

[0017] Update the strong classifier of this round to be,

[0018]

[0019] Let f0(x) be the initial model, and obtain f by iteratively combining T base models. T (x), which is the final strong classifier, the formula is,

[0020]

[0021] The internal classification and regression tree of LightGBM is used as the basic classifier. Through the cumulative model and step-up algorithm, multiple separate decision tree classifiers are combined to obtain the final strong classifier to improve the classification accuracy.

[0022] The DNN is a fully connected model composed of multiple layers of neurons, which is used to solve machine learning problems, including image classification, natural language processing and speech classification. The DNN is connected by multiple neuron layers and has a multi-layer perceptron model structure. The input data passes through each layer in turn, undergoes nonlinear transformation through the activation function, and is passed to the next layer for processing.

[0023] The DNN includes an input layer, a hidden layer and an output layer; wherein the input layer contains fault detection features, and the output layer is responsible for predicting different types of faults and fault location; for each layer l, the kth neuron output Its mathematical expression is,

[0024]

[0025] In the formula, is the weight coefficient from the kth neuron in the l-1th layer to the sth neuron in the lth layer; is the offset coefficient of the sth neuron in the lth layer; σ is the activation function; is the activation value of the i-th neuron in the l-1th layer.

[0026] The activation functions include step function, Sigmoid function and ReLU function; Shapley value analysis method is used as the interpretation framework after modeling, and each feature is calculated to predict the model The contribution of explaining the prediction of the instance is given by

[0027]

[0028] Where: g(x) is the predicted value of the model; n is the number of features; φ0 is the predicted value of the benchmark sample; φ j is the feature attributed Shapley value of feature j; when all feature values ​​exist, the sum of the feature attributions is equal to the output of the model to be explained, that is, the strong classifier output f(x) = g(x);

[0029] The marginal contribution of each feature is calculated using the cooperative game theory value allocation method to calculate the marginal contribution of each feature predicted by the model. In order to obtain the Shapley value φ of the feature j , take the average of all feature contributions to quantify the impact of each feature on model prediction, that is, perform a weighted sum on all feature value combinations; the calculation formula is,

[0030]

[0031] In the formula, S is the feature subset; |S| is the number of features in the set S; f x (S) is the model output value under the feature combination S; x j is the influence value of feature j on the model output; {x1,x2,…,x n} is a feature set of running data samples for all input active distribution network line measurement points.

[0032] In order to further identify and predict the possibility of future active distribution network failures by training neural networks, the model is processed as follows. By calculating the contribution value of each feature to the model prediction, it is convenient to improve the learning efficiency of the combined model in the process of data calculation and processing. At the same time, the Shapley value method is used to calculate the contribution value of the feature quantity to visualize the model analysis process. The calculation formula is:

[0033]

[0034] In the formula, The probability output predicted by the model;

[0035] In order to verify the recognition ability of the combined model on the existing data set and ensure that the proportion of each category in the generated training set and test set is consistent with the original data set, the following formula is used for performance evaluation:

[0036]

[0037] Where ACC is the model measurement accuracy; R is the model recall, that is, the ratio of the number of samples correctly identified as positive by the model to the total number of samples that are actually positive; TP is the number of samples correctly predicted as positive by the model; FP is the number of samples that the model incorrectly predicts as negative; FN is the number of samples that the model incorrectly predicts as positive; TN is the number of samples correctly predicted as negative.

[0038] Beneficial effects of the present invention: This method is based on the gradient boosting framework of the LightGBM (Light Gradient Boosting Machine) model, which has strong internal processing capabilities for classification features, fast classification speed and low memory usage. The fault voltage and voltage fundamental wave of the measuring point are used as feature indicators, and the voltage peak, phase-to-phase voltage difference and 6th harmonic components of each frequency band are extracted as fault feature samples. Then, the normal and fault operation data of the active distribution network are classified based on the feature relationship, wherein the fault operation data includes short-circuit fault data of different line types. Subsequently, the classification results of the fault feature samples are input as feature data into the DNN (Deep Neural Network) neural network for training and learning. Finally, the Shapley value analysis method is introduced to determine the contribution of each feature to the prediction of the combined model when faults occur in different branches of the active distribution network, thereby realizing the visualization of influencing factors in the fault diagnosis decision-making process. The present invention helps to reduce the amount of subsequent feature data collected by the model and improve the operating efficiency of line fault detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 Schematic diagram of the ReLU function of the present invention;

[0040] Figure 2It is a structural diagram of an active power distribution network simulation model of the present invention;

[0041] Figure 3 is a combined model fault detection flow chart of the present invention;

[0042] Figure 4 This is a combined model fault identification result diagram under the introduction of the signal-to-noise ratio of the present invention;

[0043] Figure 5 is a feature importance analysis diagram of the present invention;

[0044] Figure 6 This is a summary diagram of the characteristic analysis of the AB fault occurring in line L2 of the present invention;

[0045] Figure 7 It is a schematic diagram of the structure of active distribution network fault detection of the combined model integrating the Shapley value analysis method of the present invention. DETAILED DESCRIPTION

[0046] The present invention is further described in detail below in conjunction with the accompanying drawings. It should be pointed out that it is only for the purpose of more clearly describing and explaining the present invention.

[0047] like Figure 1-7 As shown, this embodiment discloses an intelligent line fault detection method for an active power distribution network, and the specific steps are as follows.

[0048] (1) The fault voltage and voltage fundamental wave at each measuring point of the active distribution network line are used as characteristic indicators to judge the line operation status;

[0049] (2) Extract the frequency band voltage peaks, phase-to-phase voltage differences, and sixth harmonic components of the line measurement points in normal and fault states as fault feature samples;

[0050] (3) Data preprocessing. Perform preprocessing operations such as denoising on the data in step 2 to improve the efficiency and accuracy of combined model training;

[0051] (4) Split the training set and test set. The data set is divided into a training set and a test set in a ratio of 7:3;

[0052] (5) Input the above classified data set into the LightGBM model for classification;

[0053] (6) Based on formula (5), multiple separate decision tree classifiers within LightGBM are combined to generate a strong classifier;

[0054] (7) Input the output value of the strong classifier into the DNN for learning and training;

[0055] (8) Fault detection performance evaluation. The thresholds of ACC and R of the model are set to 95%. If the detection results (ACC, R) output by the DNN are all lower than their corresponding thresholds, no results will be output and the process will return to step 3.

[0056] (9) Output the fault detection results. The detection results that are higher than the set threshold are taken as the final detection results and output. At the same time, the Shapley value is used to analyze the contribution of the data of different operating states of each branch of the active distribution network to the characteristics of the combined model prediction results.

[0057] The amount of operating data of different branches of the active distribution network is large and the amount of calculation is large, which leads to a long model running time, which is not conducive to the model's rapid detection of operating faults. Therefore, there is an urgent need for an algorithm model with time-saving and high-computation fault classification capabilities. LightGBM is based on the traditional Gradient Boosting Decision Tree (GBDT) algorithm and innovates. It uses the second-order Taylor series to expand the loss function, adopts the Leaf-Wise decision tree growth strategy, and deeply limits the data expansion process, significantly shortening the training time and reducing memory consumption. It can significantly reduce the risk of overfitting and improve the model's operating efficiency. The basic implementation steps of the strong classifier generation part of LightGBM are as follows.

[0058]

[0059] In the formula, r ti It is the residual calculation value, which is used to measure the current model prediction value f t-1 (x i ) and y i Deviation of x i is the input characteristic value of the data sample of the i-th active distribution network line operation measurement point; y i is the true value of the operating measurement point of the i-th active distribution network line.

[0060] Fitting is performed based on the Leaf-Wise decision tree growth strategy, using (x i ,r ti ) fits a CART and obtains the tth regression tree, whose corresponding leaf node region is R tm (m=1,2,…,M), where M is the number of leaf nodes, and for each leaf node, find the solution to minimize the loss function, that is, to fit the best output value h of the leaf node t (x).

[0061]

[0062] In the formula, h t(x) is the output value of the fitted leaf node; H is the hypothesis space of the weak learner, usually all possible CART decision trees; R tm is the region of the mth leaf node in the decision tree; h is the weak learner fitted in the current round.

[0063] Let I(x∈R tm ) is the input variable x in R tm The fitting function h of the decision tree is a simple parameterized function of the leaf node area. t (x) is as follows.

[0064]

[0065] In the formula, c tm is the output value of the mth leaf node in the decision tree.

[0066] Update the strong classifier of this round to:

[0067]

[0068] Let f0(x) be the initial model, and obtain f by iteratively combining T base models. T (x) is the final strong classifier, as shown in formula (5).

[0069]

[0070] Using the internal classification and regression tree of LightGBM as the basic classifier, multiple separate decision tree classifiers are combined to obtain the final strong classifier through the cumulative model and step-by-step improvement algorithm, which can improve the classification accuracy.

[0071] In the combined model, DNN is a fully connected model composed of multiple layers of neurons. It is a type of feedforward neural network and is usually used to solve complex machine learning problems, such as image classification, natural language processing, and speech classification. DNN is connected by multiple neuron layers and has a structure similar to the Multilayer Perceptron (MLP) model. The input data passes through each layer in turn, undergoes nonlinear transformation through the activation function, and is passed to the next layer for processing. The DNN structure in the present invention mainly includes an input layer, a hidden layer, and an output layer. Among them, the input layer contains fault detection features, and the output layer is responsible for predicting different types of faults and fault location. Therefore, for each layer l No. k Neuron output Its mathematical expression is as shown in formula (3-5).

[0072]

[0073] In the formula, is the weight coefficient from the kth neuron in the l-1th layer to the sth neuron in the lth layer; is the offset coefficient of the sth neuron in the lth layer; σ is the activation function; is the activation value of the i-th neuron in the l-1th layer. Among them, the commonly used activation functions are mainly step functions, Sigmoid functions, ReLU functions, etc. The three function images are as follows Figure 1 As shown in the figure, the ReLU function can make the network have a certain sparsity and enhance the training speed. Therefore, the present invention decides to use the ReLU function as the activation function of the DNN, as shown in formula (7).

[0074]

[0075] Where h(x) is the output of the ReLU activation function; x is the input value.

[0076] Traditional machine learning fault detection combination models usually use complex nonlinear function superposition decision-making methods, resulting in poor model interpretation ability. When performing fault detection tasks, only the results can be obtained, but no specific useful information about the feature quantity can be obtained, which makes the model uninterpretable in practical applications and has certain limitations. Therefore, it is necessary to deeply analyze the output decision of the model. This paper introduces the Shapley value analysis method as the interpretation framework after modeling, and calculates the impact of each feature on the model prediction. The contribution of exploratory examples is shown in Equation (8).

[0077]

[0078] Where: g(x) is the predicted value of the model; n is the number of features; φ0 is the predicted value of the benchmark sample; φ j is the feature attributed Shapley value of feature j. When all feature values ​​exist, the sum of feature attributions is equal to the output of the model to be explained, that is, the strong classifier output f(x) = g(x). The marginal contribution of each feature is calculated using the cooperative game theory value allocation method to calculate the marginal contribution of each feature predicted by the model. In order to obtain the Shapley value φ of the feature j , take the average of all feature contributions to quantify the impact of each feature on model prediction, that is, perform a weighted sum on all feature value combinations. The calculation formula is as follows.

[0079]

[0080] In the formula, S is the feature subset; |S| is the number of features in the set S; f x (S) is the model output value under the feature combination S; x j is the influence value of feature j on the model output; {x1,x2,…,xn} is a feature set of running data samples for all input active distribution network line measurement points.

[0081] In order to further identify and predict the possibility of future active distribution network failures by training neural networks, the model needs to be processed as follows. By calculating the contribution value of each feature to the model prediction, it is convenient to improve the learning efficiency of the combined model during data calculation and processing. At the same time, the Shapley value method is used to calculate the contribution value of the feature quantity to visualize the model analysis process. The calculation formula is shown below.

[0082]

[0083] In the formula, The probability output of the model prediction.

[0084] In order to verify the recognition ability of the combined model on the existing dataset and ensure that the proportion of each category in the generated training set and test set is consistent with the original dataset, the following formulas (12)-(13) are used for performance evaluation.

[0085]

[0086]

[0087] Where ACC is the model measurement accuracy; R is the model recall, that is, the ratio of the number of samples correctly identified as positive by the model to the total number of samples that are actually positive; TP is the number of samples correctly predicted as positive by the model; FP is the number of samples that the model incorrectly predicts as negative; FN is the number of samples that the model incorrectly predicts as positive; TN is the number of samples correctly predicted as negative.

[0088] The results of active distribution network line fault location and identification are shown in Table 1.

[0089] Table 1 Fault location and identification results

[0090]

[0091] It can be seen from Table 1 that the LightGBM-DNN combined model proposed in the present invention has excellent performance. It not only has a high recall rate, but also shows strong generalization ability. At the same time, it has excellent recognition accuracy, which provides a solid foundation for the subsequent active distribution network fault type identification.

[0092] Based on this, Figure 2 The middle line L9 and the measuring point 8 are used as observation objects to carry out different types of line fault type identification test experiments. The line fault type identification results of the combined model of the present invention are shown in Table 2.

[0093] Table 2 Fault type identification results

[0094]

[0095]

[0096] It can also be seen from Table 2 that the LightGBM-DNN combined model of the present invention still has a strong generalization ability in line fault type recognition, and the fault type recognition and classification results have a high accuracy rate.

[0097] Considering the operating conditions of active distribution networks in real environments, it is decided to introduce the signal-to-noise ratio to study the noise resistance of the combined model of the present invention. Figure 2 When an AG fault occurs on the middle line L9, a 20-40dB signal-to-noise ratio is introduced to observe the noise robustness of the combined model. The fault detection results are as follows: Figure 4 shown.

[0098] As can be seen from the figure, when the signal-to-noise ratio is between 20 and 40 dB, the recall rate of the combined model of the present invention reaches more than 95%, and with the increase of the signal-to-noise ratio, the trend of accuracy improvement is the most stable, the performance is excellent, and the noise resistance is strong, which is suitable for application in the actual active distribution network environment.

[0099] In order to analyze the influence of feature quantities on specific fault location, the Shapley value method is introduced to sort based on feature importance. Figure 5 , 6 The characteristic importance and fault characteristic analysis summary diagram of measurement point 2 when AB fault occurs on line L2 are listed. Figure 5 This is an analysis diagram of the importance of the characteristic quantity for judging the fault type of line L2. It can be seen from the figure that for the L2 line, the voltage difference between each phase at the measurement point and the zero-sequence voltage change have different contribution values ​​when the model is used to judge the line fault type. Among them, U b -U c The importance is highest when AB fault occurs on line L2. c -U a The importance is the lowest, and the importance analysis of other line fault type feature quantities can also be arranged from high to low according to the degree of contribution. Figure 6The following is a summary of the fault discrimination feature analysis when AB fault occurs. The SHAP values ​​corresponding to different feature quantities are distributed on both sides of the middle baseline. The position of the point indicates the positive and negative impact of the feature on the prediction result, as well as the magnitude of its impact. Blue represents negative impact, and red represents positive impact. Dark colors represent higher feature values, and lighter colors indicate that the feature value is closer to the middle range. The size of the scattered points represents the absolute contribution of the feature value to the model prediction. It can be seen from the figure that the second harmonic (150Hz) and the fifth harmonic (300Hz) have a greater impact on the fault location model. The introduction of the Shapley value analysis method realizes the visualization of the contribution values ​​of different feature quantities in the process of active distribution network line fault detection, which is convenient for the collection of fault feature quantities of different lines in the active distribution network in the future, helps the attribution analysis of fault diagnosis results, reduces the amount of data required for model operation, improves the credibility of diagnosis results, and provides effective detection information for the fault detection area of ​​the active distribution network.

[0100] In order to verify the practical applicability of the present invention, a simulation model of an IEEE 33-node active distribution network is constructed based on Matlab / Simulink software. Figure 2 As shown. The measurement points of different branches of the distribution network numbered 9, 22, 24, and 33 correspond to four distributed power sources. Among them, DG1 is a grid-connected photovoltaic power station, DG2 and DG4 are grid-connected inverter direct-drive wind turbines, and DG3 is a user-provided power supply. At the same time, in order to enhance the diversity of simulation samples, it is decided to use the phase A voltage of the measurement point in the constructed active distribution network model as the reference point, and simulate four different fault types at 40-50% of the node line, namely: single-phase ground fault (AG), AB phase short circuit (AB), AB two-phase ground short circuit (AB-G) and three-phase short circuit (ABC).

[0101] The implementation of the overall structural framework of the combined model of the present invention is mainly divided into the following two stages: Figure 7 shown.

[0102] (1) The active distribution network operation feature processing and screening process based on LightGBM-DNN specifically includes the following three modules.

[0103] Module 1: LightGBM-based classification of preliminary data features of active distribution network operation. The original active distribution network operation data is loaded into the internal gradient boosting framework, and preliminary high-quality feature screening operations are performed on the data;

[0104] Module 2: Data feature training and learning phase based on DNN. DNN takes the data features based on the LightGBM model as input and uses its own active learning advantages to learn and train the data features of the active distribution network operation;

[0105] (2) Model result evaluation based on Shapley value. The following is the module content.

[0106] Module 3: Based on the Shapley value, the characteristic contribution of the data of different operating states of each branch of the active distribution network to the prediction results of the LightGBM-DNN combined model is calculated to improve the transparency of the model fault detection and decision-making process. On this basis, the Shapley value analysis results are re-input into Module 1 as feedback to facilitate the guidance of the model's subsequent active distribution network data collection work, while visualizing the characteristic contribution of short-circuit faults of different lines in the distribution network and reducing the amount of subsequent data collection for the model.

[0107] The structural schematic diagram of the present invention is as follows Figure 7 As shown in the figure. First, the fault voltage and voltage fundamental wave at each measuring point of the active distribution network line are used as characteristic indicators to judge the operating status of the line. Then, the frequency band voltage peak, phase-to-phase voltage difference and harmonic components of the normal and faulty states of the line measurement points are extracted as fault feature samples. Next, the collected data set is input into the LightGBM model for classification, and the output value of the LightGBM strong classifier is input into the DNN for learning and training. At the same time, the fault detection performance evaluation is set. If each detection result output by the combined model is lower than its corresponding threshold, the result will not be output and the detection will be restarted. Finally, the detection result higher than the set threshold is taken as the final detection result and the Shapley value is used for structural contribution value analysis.

[0108] This embodiment classifies the data of the normal operation of the active distribution network and the data of the fault operation based on the above-mentioned characteristic relationship, wherein the fault operation data covers the short-circuit fault conditions of different lines. Then, the classification results of the fault feature samples are input into the neural network as feature data for training and learning. Finally, the Shapley value analysis method is introduced into the model to determine the contribution of each fault feature to the prediction of the combined model when faults occur in different branches of the active distribution network, thereby realizing the visualization of the influencing factors in the fault diagnosis decision. This method not only helps to reduce the amount of feature data collection required for the model, but also improves the operating efficiency of the model. Therefore, the present invention uses artificial intelligence technology based on the characteristic relationship between voltage and line fault to realize fault detection of the active distribution network line, thereby effectively promoting the intelligent and efficient operation of the active distribution network. Solve the problems of complex system topology and difficult line fault location caused by the access of a large number of distributed power sources to the distribution network. First, based on the gradient boosting framework of LightGBM (Light Gradient Boosting Machine, LightGBM) model, which has strong internal processing ability of classification features, fast classification speed and low memory usage, the method uses the fault voltage and voltage fundamental of the measurement point as feature indicators, and extracts the voltage peak, phase-to-phase voltage difference and 6th harmonic components of each frequency band as fault feature samples. Then, the normal and fault operation data of the active distribution network are classified based on the feature relationship, where the fault operation data includes short-circuit fault data of different line types. Subsequently, the classification results of the fault feature samples are input as feature data into the DNN (Deep Neural Network, DNN) neural network for training and learning. Finally, the Shapley value analysis method is introduced to determine the contribution of each feature to the prediction of the combined model when faults occur in different branches of the active distribution network, so as to realize the visualization of the influencing factors in the fault diagnosis decision-making process. The present invention helps to reduce the amount of subsequent feature data collection of the model and improve the efficiency of line fault detection operation.

[0109] The above are only preferred specific implementation modes of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical solutions and concepts of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for detecting intelligent line faults in an active power distribution network, characterized in that: The following steps are involved: S1. The fault voltage and voltage fundamental wave at each measuring point of the active distribution network line are used as characteristic indicators for judging the line operation status; S2, extracting the frequency band voltage peak value, phase-to-phase voltage difference and 6th harmonic component of the line measurement point in normal and fault states as fault feature samples; S3, performing data preprocessing, including performing a denoising pre-operation on the data in S2; S4, split the training set and the test set, and divide the data set into a training set and a test set in a ratio of 7:3; S5, inputting the classified data set into the model for classification; S6. Combine multiple separate decision tree classifiers within LightGBM to generate a strong classifier; S7, input the output value of the strong classifier into the DNN for learning and training; S8. Perform fault detection performance evaluation; S9. Output fault detection results.

2. The method for detecting intelligent line faults in an active power distribution network according to claim 1, characterized in that: The LightGBM model includes extending the loss function through the second-order Taylor series, adopting the Leaf-Wise decision tree growth strategy, deeply limiting the data expansion process, shortening the training time, and reducing memory consumption, which is used to reduce the risk of overfitting and improve the model operation efficiency; the implementation steps of the strong classifier generation part of LightGBM are as follows, In the formula, r ti It is the residual calculation value, which is used to measure the current model prediction value f t-1 (x i ) and y i Deviation of x i is the input characteristic value of the data sample of the i-th active distribution network line operation measurement point; y i is the true value of the operating measurement point of the i-th active distribution network line; Fitting is performed based on the Leaf-Wise decision tree growth strategy, using (x i ,r ti ) fits a CART and obtains the tth regression tree, whose corresponding leaf node region is R tm (m=1,2,…,M), where M is the number of leaf nodes, and for each leaf node, find the solution to minimize the loss function, that is, to fit the best output value h of the leaf node t (x), In the formula, h t (x) is the output value of the fitted leaf node; H is the hypothesis space of the weak learner, usually all possible CART decision trees; R tm is the region of the mth leaf node in the decision tree; h is the weak learner fitted in the current round; Let I(x∈R tm ) is the input variable x in R tm The fitting function h of the decision tree is a simple parameterized function of the leaf node area. t (x) is, In the formula, c tm is the output value of the mth leaf node in the decision tree.

3. The method for detecting intelligent line faults in an active power distribution network according to claim 2, characterized in that: Update the strong classifier of this round to be, Let f0(x) be the initial model, and obtain f by iteratively combining T base models. T (x), which is the final strong classifier, the formula is, The internal classification and regression tree of LightGBM is used as the basic classifier. Through the cumulative model and step-up algorithm, multiple separate decision tree classifiers are combined to obtain the final strong classifier to improve the classification accuracy.

4. The method for detecting intelligent line faults in an active power distribution network according to claim 1, characterized in that: The DNN is a fully connected model composed of multiple layers of neurons, which is used to solve machine learning problems, including image classification, natural language processing and speech classification. The DNN is connected by multiple neuron layers and has a multi-layer perceptron model structure. The input data passes through each layer in turn, undergoes nonlinear transformation through the activation function, and is passed to the next layer for processing.

5. The method for detecting intelligent line faults in an active power distribution network according to claim 4, characterized in that: The DNN includes an input layer, a hidden layer and an output layer; wherein the input layer contains fault detection features, and the output layer is responsible for predicting different types of faults and fault location; for each layer l, the kth neuron output Its mathematical expression is: In the formula, is the weight coefficient from the kth neuron in the l-1th layer to the sth neuron in the lth layer; is the offset coefficient of the sth neuron in the lth layer; σ is the activation function; is the activation value of the i-th neuron in the l-1th layer.

6. The method for detecting intelligent line faults in an active power distribution network according to claim 5, characterized in that: The activation functions include step function, Sigmoid function and ReLU function; Shapley value analysis method is used as the interpretation framework after modeling, and each feature is calculated to predict the model The contribution of explaining the prediction of the instance is given by Where: g(x) is the predicted value of the model; n is the number of features; φ0 is the predicted value of the benchmark sample; φ j is the feature attributed Shapley value of feature j; when all feature values ​​exist, the sum of the feature attributions is equal to the output of the model to be explained, that is, the strong classifier output f(x) = g(x); The marginal contribution of each feature is calculated using the cooperative game theory value allocation method to calculate the marginal contribution of each feature predicted by the model. In order to obtain the Shapley value φ of the feature j , take the average of all feature contributions to quantify the impact of each feature on model prediction, that is, perform a weighted sum on all feature value combinations; the calculation formula is, In the formula, S is the feature subset; |S| is the number of features in the set S; f x (S) is the model output value under the feature combination S; x j is the influence value of feature j on the model output; {x1,x2,…,x n } is a feature set of running data samples for all input active distribution network line measurement points.

7. The active power distribution network intelligent line fault detection method according to claim 6, characterized in that: In order to further identify and predict the possibility of future active distribution network failures by training neural networks, the model is processed as follows. By calculating the contribution value of each feature to the model prediction, it is convenient to improve the learning efficiency of the combined model in the process of data calculation and processing. At the same time, the Shapley value method is used to calculate the contribution value of the feature quantity to visualize the model analysis process. The calculation formula is: In the formula, The probability output predicted by the model; In order to verify the recognition ability of the combined model on the existing data set and ensure that the proportion of each category in the generated training set and test set is consistent with the original data set, the following formula is used for performance evaluation: Where ACC is the model measurement accuracy; R is the model recall, that is, the ratio of the number of samples correctly identified as positive by the model to the total number of samples that are actually positive; TP is the number of samples correctly predicted as positive by the model; FP is the number of samples that the model incorrectly predicts as negative; FN is the number of samples that the model incorrectly predicts as positive; TN is the number of samples correctly predicted as negative.

Images