Power distribution network high-resistance fault detection and interpretability analysis method

Abstract

The invention provides a power distribution network high-resistance fault detection and interpretability analysis method, and belongs to the technical field of power distribution network fault diagnosis and artificial intelligence deep learning. According to the method, an original transient zero-sequence current signal is decomposed and reconstructed through improved adaptive noise complete set empirical mode decomposition, and a reconstructed transient zero-sequence current signal is generated; constructing a time convolution network model based on the reconstructed transient zero-sequence current signal, and outputting a fault detection result; and further constructing a score weighted class activation mapping method to carry out qualitative analysis and quantitative calculation on the detection basis of the model, and carrying out closed-loop feedback on the interpretability of a data-driven fault detection scheme by depicting the matching degree between the high-resistance zero-break characteristic and a model decision concern area. The power distribution network high-resistance fault detection and interpretability analysis method has high detection accuracy and interpretability under different operation conditions, and is suitable for high-resistance fault detection and analysis in a complex power distribution network environment.

Description

Technical Field

[0001] The present invention relates to the technical field of distribution network fault diagnosis and artificial intelligence deep learning, and particularly relates to a method for detecting high-resistance faults and interpretable analysis in a distribution network. Background Art

[0002] There are numerous lines, complex structures, and relatively close distances to the ground in a distribution system, making it easy to form high-resistance grounding faults through non-metallic conductive media such as grasslands, concrete, and tree branches. Due to the large transition resistance of high-resistance faults and susceptibility to harmonic interference, their fault characteristics are weak and similar to disturbance signals caused by conventional switching events such as load switching, capacitor switching, and magnetizing inrush current, making it difficult to be effectively detected and disposed of in a timely manner. If high-resistance faults persist for a long time, they may damage equipment, cause fires, or even further expand the fault range, resulting in serious losses. Therefore, studying sensitive and reliable high-resistance fault detection schemes is of great significance for ensuring the safe and reliable operation of the distribution system.

[0003] Currently, according to the differences in the feature analysis dimension of the high-resistance grounding fault detection scheme for a distribution network, it can be mainly divided into two categories: one is the index threshold method based on electrical quantity features, and the other is the artificial intelligence method based on data-driven. The index threshold method quantifies and analyzes the differences in electrical information such as voltage and current before and after a fault in the time domain, frequency domain, and time-frequency domain, and sets feature thresholds to detect high-resistance grounding faults. However, the above methods all rely on single feature quantities or local features to construct criteria, and their thresholds are usually set by artificial experience. In complex operation scenarios such as increasingly complex system structures and measurement noise interference, classification blind spots are likely to occur, and their generality needs to be further expanded. In recent years, the continuous development of data-driven artificial intelligence technology has provided new research ideas for solving the problem of high-resistance fault detection. The artificial intelligence method does not rely on complex mechanism analysis and can fit the non-linear mapping relationship between input samples and output results from a large amount of data, thereby achieving fast and accurate fault detection. Some scholars have applied artificial intelligence to high-resistance fault detection, such as support vector machines, artificial neural networks, convolutional neural networks, etc.

[0004] However, existing methods generally focus on the application of artificial intelligence algorithms, and the models are similar to "black boxes", having problems such as low confidence and weak decision interpretable basis, resulting in risks when artificial intelligence algorithms are used to complete safety-sensitive tasks such as distribution network fault diagnosis. At the same time, the quantitative analysis and evaluation of power operation and maintenance knowledge from the perspective of interpretability have not been carried out, and the performance of multiple quantitative indicators has not been comprehensively considered to evaluate the effectiveness and reliability of model decisions. Generally speaking, although the research on the interpretability of power artificial intelligence has been preliminarily attempted in many fields, it is still in its infancy in the field of distribution network fault diagnosis. Summary of the Invention

[0005] The object of the present invention is to provide a method for high-resistance fault detection and interpretable analysis of a distribution network, so as to overcome the problems in the background art, and the effectiveness of the proposed solution and its applicability in real scenarios have strong application value.

[0006] To achieve the above object, the present invention provides a method for high-resistance fault detection and interpretable analysis of a distribution network, including the following steps:

[0007] Step S1: Decompose and reconstruct the original transient zero-sequence current signal by using improved adaptive noise complete ensemble empirical mode decomposition, and generate a reconstructed transient zero-sequence current signal;

[0008] Step S2: Based on the reconstructed transient zero-sequence current signal generated in Step S1, construct a temporal convolutional network model and output a fault detection result;

[0009] Step S3: Based on the fault detection result output in Step S2, use the fractional weighted class activation mapping method to generate an attribution heat map, analyze the correlation between the reconstructed transient zero-sequence current signal and the fault detection result, and construct a quantitative evaluation index;

[0010] Step S31: The class activation mapping method generates a class activation mapping graph by linearly weighting and fusing the feature fusion weight and the feature map;

[0011] Step S32: Calculate the attribution values of each sampling point in the time series of the reconstructed transient zero-sequence current signal;

[0012] Step S33: Further construct a quantitative evaluation index based on the attribution values generated in Step S32.

[0013] Preferably, Step S1 specifically includes:

[0014] Step S11: Add Gaussian white noise to the original transient zero-sequence current signal, and calculate the residual value and modal component of the first decomposition;

[0015] Step S12: Repeat the step of superimposing Gaussian white noise until no further decomposition is possible, and decompose the original transient zero-sequence current signal into the sum of multiple intrinsic mode function IMF components and a residual;

[0016] Select the components above IMF5 for signal superposition to form a reconstructed transient zero-sequence current signal.

[0017] Preferably, the temporal convolutional network model includes an input layer, multiple TCN modules, a 1×1 convolutional layer, a Flatten layer, a Dense layer, and a Softmax layer connected in sequence.

[0018] Preferably, each TCN module includes: a causal dilated convolutional layer, a weight normalization layer, a ReLU activation function, and a Dropout unit.

[0019] Preferably, the method for calculating the class activation mapping diagram is as follows:

[0020]

[0021] where L represents the class activation mapping diagram; ReLU represents the activation function; represents the channel weight; k represents the index of the channel; c represents the category of interest; represents the activation output of the k-th channel of the l-th convolutional layer in the model.

[0022] Preferably, the steps for constructing the quantization evaluation index are as follows:

[0023] For the zero-sequence current waveform containing T sampling points, define the signal zero-crossing moment as t0, and take the time sequence interval [t0 - ΔT, t0 + ΔT] of ΔT sampling points before and after this moment as the analysis window, and define it as the waveform zero-crossing key area set Ω1:

[0024] Ω1 = {t|t ∈ [t0 - ΔT, t0 + ΔT]};

[0025] Calculate the attribution value S(t) ∈ [0, 1] of the feature at the sampling point moment t in the transient zero-sequence current time series, and define the high-attribution area set Ω2:

[0026] Ω2 = {t|S(t) > τ};

[0027] where τ represents the attribution value threshold;

[0028] Define the matching degree index ZAM between the high-attribution area and the key area, and the global proportion index KAR of the key area attribution:

[0029]

[0030] where |·| represents the number of elements contained in the set.

[0031] Therefore, the present invention adopts the above-mentioned method for detecting high-resistance faults and interpretable analysis in the distribution network, and the beneficial technical effects are as follows:

[0032] The present invention can perform high-precision high-resistance fault detection in the complex operating environment of a new distribution system and conduct visual analysis on the model decision-making mechanism. On the one hand, the visual results intuitively present the concerned segments of the model for the zero-sequence current sequence, providing qualitative and tuning guidance for hyperparameter selection. On the other hand, combined with quantitative evaluation indicators, it explains the degree of dependence of the model on the waveform distortion during the zero-rest period in the detection results, enhancing the interpretability of the model decision-making process. The research results can provide technical support for the application of high-resistance fault detection methods based on deep learning in actual systems. Brief Description of the Drawings

[0033] Figure 1 is the topology of a 10 kV distribution network;

[0034] Figure 2 is the Emanuel model;

[0035] Figure 3 is the flow chart of the high-resistance fault detection and interpretability analysis method for the distribution network;

[0036] Figure 4 is the performance on the training set and the test set during the model training process;

[0037] Figure 5 is the confusion matrix of the model on the training set and the test set; among them, Figure 5 in (a) is the confusion matrix on the training set; Figure 5 in (b) is the confusion matrix on the test set;

[0038] Figure 6 is the t-SNE visualization result; among them, Figure 6 in (a) is the visualization distribution result of the original data set after dimensionality reduction by the t-SNE algorithm; Figure 6 in (b) is the visualization distribution result of the data set processed by 1 TCN module after dimensionality reduction by the t-SNE algorithm; Figure 6 in (c) is the visualization distribution result of the data set processed by 2 TCN modules after dimensionality reduction by the t-SNE algorithm; Figure 6 in (d) is the visualization distribution result of the data set processed by 3 TCN modules after dimensionality reduction by the t-SNE algorithm;

[0039] Figure 7 is to use Score-CAM to analyze the decision-making mechanism of samples under different operating conditions; among them, Figure 7 in (a) is the attribution heat map of high-resistance fault samples; Figure 7 in (b) is the attribution heat map of capacitor switching samples; Figure 7 in (c) is the attribution heat map of load switching samples; Figure 7 in (d) is the attribution heat map of inrush current samples;

[0040] Figure 8 serve as the decision basis for different convolution kernel sizes; among them, Figure 8 in (a) is the attribution heatmap when the convolution kernel size is 1; Figure 8 in (b) is the attribution heatmap when the convolution kernel size is 3; Figure 8 in (c) is the attribution heatmap when the convolution kernel size is 5;

[0041] Figure 9 is the residual block structure in the TCN. Specific implementation manner

[0042] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0043] Unless otherwise defined, the technical terms or scientific terms used in the present invention shall have the ordinary meanings understood by those of ordinary skill in the field to which the present invention belongs.

[0044] Embodiment 1

[0045] As Figure 3 shown, the flowchart of the high-resistance fault detection and interpretability analysis method for the distribution network includes four stages: data preprocessing, model training and parameter tuning, actual application, and interpretability analysis.

[0046] Data preprocessing stage: Use a measurement device to collect the original transient zero-sequence current signals under high-resistance fault and normal disturbance conditions, add random noise to them to simulate the complex operating environment of the distribution system, and decompose and reconstruct the noisy original transient zero-sequence current signals via the Improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) algorithm. Considering the uncertainty of the actual system fault occurrence time and the action delay of the recorder startup, the reconstructed transient zero-sequence current signals are processed using a sliding window segmentation (the time window used in this embodiment is 40 ms), and the segmented transient zero-sequence current signal x is normalized using Equation (1):

[0047]

[0048] where x max and x min respectively represent the maximum and minimum values in the transient zero-sequence current signal x before normalization; x i and x iLet \(x_i\) and \(x_i'\) represent the \(i\)-th element values in the signals \(x\) and \(x'\) before and after normalization, respectively. Through the above operations, the order-of-magnitude differences of electrical quantities can be reduced, enabling the model to pay more attention to the data change trends and improving the model performance.

[0049] Different from the traditional empirical mode decomposition (EMD) method, ICEEMDAN improves the stability and accuracy of the decomposition process by optimizing the noise addition and decomposition processes. On the basis of effectively solving mode mixing, it can reduce the residual noise of the intrinsic mode components and further highlight the time and frequency domain characteristics of each component. The present invention uses the ICEEMDAN method to decompose and reconstruct the noisy transient zero-sequence current signal, and the steps are as follows:

[0050] Step S11: Define the transient zero-sequence current signal to be decomposed as \(i_0\), \(E\) K (·) represents the \(K\)-th order intrinsic mode function component (IMF) after EMD decomposition. Add Gaussian white noise to \(i_0\), and calculate the residual value \(R_1\) of the first decomposition and the first mode component \(D_1\):

[0051]

[0052] \(D_1 = i_0 - R_1\) (4);

[0053] where \(\varepsilon\) represents the signal-to-noise ratio of Gaussian white noise, \(w\) (n) represents the \(n\)-th group of added Gaussian white noise, and \(N\) represents the total number of additions.

[0054] Step S12: Repeat the step of superimposing noise in Step S11, and calculate the residual value \(R\) j and the mode component \(D\) j after the \(j\)-th decomposition until it cannot be decomposed further. Through the above steps, the transient zero-sequence current signal \(i_0\) to be decomposed is decomposed into the sum of multiple IMF components and residuals.

[0055] By performing spectral analysis on each decomposed IMF component, it can be seen that the signal characteristics contained in some IMF components are measurement noise characteristics and have a weak correlation with the original transient zero-sequence current signal. Therefore, the present invention selects the components above IMF5 for signal superposition to form a reconstructed signal, which can effectively filter out noise interference while fully retaining the fault characteristics.

[0056] Model training and parameter tuning stage: Divide the preprocessed dataset into a training set and a test set in a ratio of 4:1; construct a temporal convolutional network model, input the training set samples into the model, and carry out model training and hyperparameter tuning. Among them, cross-entropy is used as the loss function and minimized through the Adam optimizer. When the loss function and accuracy tend to be stable, save the model weights for subsequent test calls and application analysis.

[0057] As Figure 9 shown, the temporal convolutional network model includes the following modules connected in sequence:

[0058] Input layer, used to receive the reconstructed transient zero-sequence current signal;

[0059] Multiple TCN modules;

[0060] 1×1 convolutional layer, used to adjust the feature dimension;

[0061] Flatten layer, used to flatten multi-dimensional features;

[0062] Dense layer, used for fully connected operations;

[0063] Softmax layer, used to output classification probabilities

[0064] Each TCN module includes:

[0065] Causal dilated convolutional layer, used to extract time series features;

[0066] Weight normalization layer, used to stabilize the training process;

[0067] ReLU activation function, used to introduce non-linearity;

[0068] Dropout unit, used to prevent overfitting.

[0069] Actual application stage: Collect the transient zero-sequence current signals to be detected, after data preprocessing, input them into the trained and saved model, and output the classification results.

[0070] Interpretability analysis stage: Use the Score-CAM method to analyze the model decision results, calculate the attribution values of the features at each sampling point moment in the transient zero-sequence current time series, and further clarify which parts of the time series dominate or affect the sample classification results. Combine with visualization methods to explain the influence of the model decision basis and hyperparameter settings on the classification results, so as to intuitively reflect the internal relationship between the input features and the output categories.

[0071] In the actual application process, reasonable thresholds can be set for ZAM and KAR to quantify the degree of attention of the model to the zero-rest distortion characteristics of high-resistance faults and evaluate the reliability of the model classification results. By comparing whether the output results of the model are consistent with the attribution features of the interpretability analysis, the controllability and credibility of the fault detection process are enhanced. Specifically, if the classification result output by the model is a high-resistance fault, but the indicators ZAM and KAR of the sample are both lower than the threshold setting, it indicates that the decision of the model does not mainly rely on the features of the key area, indicating that it pays insufficient attention to the core features of high-resistance faults and there may be a risk of misjudgment; if the discrimination result output by the model is not a high-resistance fault, but its indicator is higher than the threshold setting, it means that the model has accurately focused on the zero-point distortion feature, but the final classification result is not determined as a high-resistance fault, which may mean a risk of missed judgment. In the above situations, an artificial intervention mechanism can be triggered to arrange technicians to conduct an in-depth review of the entire process of fault diagnosis, further analyze the discrimination basis of the model, and clarify the source of potential problems. In this way, the risks of misjudgment and missed judgment are effectively reduced, and the accuracy and credibility of high-resistance fault detection are improved.

[0072] Score-CAM generates a class activation mapping diagram by linearly weighting and fusing the feature fusion weight and the feature map to achieve a more stable interpretation effect. For a given temporal convolutional network model, let its input and output be x and Y respectively, satisfying Y = f(x). Select the k-th channel of the l-th convolutional layer from it, and the corresponding activation is For the known input x i , The contribution to the output Y Can be defined as:

[0073]

[0074] Among them, Indicates the Hadamard product; Is a vector with the same shape as x i The definition formula is:

[0075]

[0076] Among them, the function s[.] represents the normalization operation, mapping each element to the interval [0, 1]; the function Up(.) represents upsampling .

[0077] Select the category c of interest, then the calculation formula of the class activation map of Score-CAM can be defined as:

[0078]

[0079] Among them, L represents the class activation mapping diagram; ReLU represents the activation function; represents the channel weight; k represents the index of the channel; represents the activation output of the k-th channel of the l-th convolutional layer in the model;

[0080] the weights of each channel can determine the specific category information contained in the class activation map; ReLU(·) is an activation function used to remove neurons that are useless for the category c of interest.

[0081] The present invention uses Score-CAM to analyze the decision results of the fault detection model based on the temporal convolutional network, and calculates the attribution value of the features at each sampling point in the transient zero-sequence current time series according to equations (5)-(7). Thus, the correlation between the input sample and the detection result is analyzed.

[0082] The class activation map generated by Score-CAM can provide a stable visualization effect, showing the key regions that the model focuses on during the classification decision process, thereby providing guidance for model hyperparameter tuning. On the basis of this qualitative analysis, the present invention further constructs a quantitative evaluation index based on the generated attribution result values, so as to provide more intuitive decision support for the operation and maintenance personnel.

[0083] First, for the transient zero-sequence current waveform containing T sampling points, define the signal zero-crossing moment (take the time point of the current sign change between adjacent sampling points) as t0, and take the time series interval [t0 - ΔT, t0 + ΔT] of ΔT sampling points before and after this moment as the analysis window, and define it as the key region set Ω1 of the waveform zero-crossing as shown in equation (8):

[0084] Ω1 = {t|t ∈ [t0 - ΔT, t0 + ΔT]} (8);

[0085] Calculate the attribution value S(t) ∈ [0, 1] of the features at the sampling point moment t in the transient zero-sequence current time series according to equations (5)-(7), and define the high-attribution region set Ω2 as shown in equation (9):

[0086] Ω2 = {t|S(t) > τ} (9);

[0087] where τ is the attribution value threshold, and in this embodiment, τ is taken as 0.6.

[0088] Further define the matching degree index (ZAM) of the high-attribution region and the key region, and the global proportion index (KAR) of the key region attribution, as shown in equations (10) and (11) respectively:

[0089]

[0090] where |·| represents the number of elements contained in the set.

[0091] By analyzing equations (10) and (11), it can be seen that ZAM measures whether the model accurately focuses on the distortion characteristics near the waveform zero-crossing point by evaluating whether the high-attribution-value points are concentrated in the key area. The larger the value of ZAM, the more the high-contribution attribution-value points of the model are mainly distributed in the key area, indicating that the model extracts more discriminative features in this area; conversely, if the value of ZAM is low, it means that the high-contribution attribution-value points of the model are more scattered, and it may rely on information in other areas for classification, reducing the dependence on the key features of the zero-crossing point. KAR reflects the proportion of the attribution value of the key area in the global attribution value, measuring the degree of dependence of the model on the key area of the zero-crossing point in the overall decision-making process. The larger the value of KAR, the more the model preferentially relies on the waveform features of the key area of the zero-crossing point rather than the noise or redundant information in other non-key areas in the classification decision, thus reflecting the global importance of the features of the key area to the model.

[0092] Overall, KAR reflects the proportion of the features of the key area in the overall attribution value, measuring the global attention preference of the model, while ZAM reflects whether the model accurately focuses on the high-contribution points in the key area, measuring its local feature capture ability. The combination of the two can effectively evaluate the attribution feature distribution and its decision-making basis of the model in high-resistance fault detection.

[0093] The following is a further illustration of the present invention through simulation experiments.

[0094] Build a 10 kV distribution network simulation model in MATLAB / Simulink as Figure 1 shown to simulate high-resistance faults and disturbances and conduct subsequent analysis. The model includes a power source G, a transformer (using the DYn11 connection method, where D represents the delta connection on the high-voltage side and Y represents the star connection on the low-voltage side), lines (L1 to L 10 ), loads (connected to the grid through transformers), and distributed power sources (DG1 to DG3). Among them: the transformer capacity is 250 MVA, the transformation ratio is 110 kV / 10.5 kV, and its neutral point is grounded through an arc suppression coil (R and L represent the resistance and inductance parameters of the arc suppression coil respectively, with a compensation degree of 8%); the lengths of each line are marked in the figure, and the line parameters are shown in Table 1. At the monitoring point, a μPMU device with a sampling frequency of 10 kHz is used to obtain the transient zero-sequence current signal, and the sampling frequency is 10 kHz.

[0095] Table 1 Line parameters

[0096]

[0097] Based on the above distribution system simulation, a simulation sample library can be established, and the parameters are shown in Table 2. Among them, the high-resistance fault is simulated using the Emanuel model as Figure 2 shown. Figure 2 In it, U pand U n is a DC voltage source with a ±10% fluctuation set to simulate the asymmetry and non-linearity of the arc voltage and fault current; R p and R n are time-varying resistors used to simulate the fault arc resistance; D p and D n are ideal diodes that, together with U p and U n collectively form the positive and negative half-cycle current paths of the circuit.

[0098] Table 2 Sample parameters

[0099]

[0100] By changing the access position, initial phase angle, and sample parameters, simulation samples occurring under different transition resistances, different line types, and positions can be fully obtained. Uniformly intercept the data within a 0.2 s time window before and after the switch event (2 cycles before the fault and 8 cycles after the fault), and segment it through a 40 ms time window. After decomposition and reconstruction using the ICEEMDAN method, labels 0, 1, and 2 are assigned according to the simulation settings, corresponding to normal, disturbance (including capacitor switching, load switching, inrush current), and high-resistance fault respectively. Finally, the dataset is divided into a training set and a test set at a ratio of 4:1. Considering that there is a certain imbalance property between the training and test samples, it is obviously biased to use only the overall accuracy as an indicator to evaluate the model performance. Therefore, precision, recall, and F1-score are supplemented to comprehensively evaluate the model performance, and the F1-score is used to evaluate the classification effect of the fault location model.

[0101] Through Figure 4 show the performance of the model on the training set and the test set during the model training process. The abscissa is the number of iterations, totaling 50 rounds; the ordinate is the F1-score and the loss respectively. From Figure 4 it can be observed that the loss curve drops significantly in the initial stage of training and then tends to be stable when the iteration reaches 40 rounds. To ensure the stable effect of the model, in this embodiment, the model saved at the 48th iteration is selected as the final model to carry out subsequent test work. The confusion matrices of the model on the training set and the test set are exported as shown in Figure 5 . It can be seen that the overall accuracy of the model on the test set reaches 96%, the precision reaches 94.75%, the recall reaches 96.01%, and the F1-score reaches 95.32%. This shows that the model can maintain a high level of generalization ability under different operating conditions and does not misjudge or miss judge the high-resistance fault category, preliminarily verifying the effectiveness of the high-resistance fault detection scheme proposed in the present invention.

[0102] To further verify the effectiveness of the proposed scheme, the t-distributed stochastic neighbor embedding algorithm is used to reduce the dimension and visualize the original data and the data processed by different TCN modules. The results are as Figure 6 shown. Different colors represent different categories; the coordinate axes only represent the distribution of data points on the two-dimensional plane after data dimension reduction, without dimension. Figure 6 As shown in (a) of Figure 6 , the original samples show considerable disorder both in terms of feature dimensions and distribution patterns, being randomly scattered in the low-dimensional visualization space and lacking an obvious aggregation trend. However, as the depth of the feature extraction layer increases (as shown in (b)-(d) of Figure 6 Figure 6 ), the sample points begin to show an aggregation trend. The distance between sample points of the same class gradually decreases, forming an obvious clustering structure. At the same time, the distance between different classes increases significantly, gradually establishing a clear classification boundary. The above transformation process fully shows that the clustering effect of the samples is gradually strengthened, indicating that the model designed in the present invention can fully mine the implicit features in the transient zero-sequence current time series and shows superior feature extraction and classification capabilities in the fault detection task.

[0103] The proposed method is compared with common fault detection models to fully verify the advantages of the proposed scheme, including two aspects: the classification effect of the model and the test time. Considering the certain randomness in the model training and testing process, the model is repeatedly tested multiple times and the mean value is calculated. The results are shown in Table 3. Among them, the test time is the average calculation time of the test set samples. The results show that the proposed method has substantial advantages in terms of computational efficiency and accuracy compared with the methods based on SVM and ANN. In addition, although CNN shows better performance in terms of test time, the TCN model shows significant superiority in terms of classification effect. It should be noted that when considering the safe and reliable operation of the actual distribution network, the consequences of missed detection and misdetection are more serious than those of delayed detection. Therefore, it is still reasonable to sacrifice a certain amount of test time to improve the accuracy of fault detection.

[0104] Table 3 Comparison of the effects of different classification methods

[0105] Solution Accuracy Precision Recall F1 Score Testing Time SVM 90.86％ 93.09％ 85.30％ 87.41％ 1.5 ms ANN 88.29％ 93.20％ 80.48％ 82.41％ 3 ms CNN 91.14％ 94.59％ 85.24％ 87.58％ 0.6 ms The method of the present invention 96.00％ 94.75％ 96.01％ 95.32％ 0.9 ms

[0106] Analyze the operating mechanism of the temporal convolutional network using the Score-CAM algorithm, that is, for the reconstructed transient zero-sequence current signal, calculate the contribution degree of each time series segment to the classification result, and construct an attribution heat map based on this, and then clearly identify the characteristic segments that play a key role in the decision result, so as to help the operation and maintenance personnel understand the decision-making basis of the model. In the attribution heat map, the attribution value ranges from (0, 1). A larger attribution value means that the sampling point has a more significant impact on the model decision, and its size is represented by different colors. According to this criterion, the influence of each segment of the transient zero-sequence current time series on the model decision result can be qualitatively evaluated. On this basis, combined with ZAM and KAR for quantitative analysis, to quantitatively evaluate the degree of attention of the model to the waveform zero-crossing interval, so as to analyze the differences in the attribution patterns of different types of samples.

[0107] Figure 7 Shows the operating mechanism of the TCN model for fault classification through Score-CAM. It can be seen that the key factor affecting the model classification result is not the peak or trough region, but the different degrees of distortion near the waveform zero-crossing point. In Figure 7 Under the high-resistance fault condition shown in (a), when the transient zero-sequence current shows a horizontal trend for a period of time after the zero-crossing point and returns to sinusoidal change, this zero-rest characteristic is effectively detected by the model and used as the key basis for high-resistance fault detection. Thus, it can be seen that the TCN model can obtain the key waveform characteristics of different types of samples, providing a visual decision-making basis for the model to make corresponding category judgments.

[0108] From the perspective of quantitative analysis, for Figure 7 the high-resistance fault samples shown in (a), the values of the indicators ZAM and KAR are 0.69 and 0.7 respectively; Figure 7 for the capacitor switching samples shown in (b), the values of the indicators ZAM and KAR are 0.61 and 0.61 respectively; Figure 7 for the load switching samples shown in (c), the values of the indicators ZAM and KAR are 0.59 and 0.63 respectively; Figure 7 for the inrush current samples shown in (d), the values of the indicators ZAM and KAR are 0.58 and 0.56 respectively. Thus, it can be seen that the indicators of the high-resistance fault samples are significantly higher than those of other disturbance type samples, indicating that the high contribution attribution values of the model are mainly concentrated in the waveform zero-crossing region, which is consistent with the electrical characteristics of high-resistance faults, indicating that the attribution results of Score-CAM have certain physical interpretability. In contrast, the two indicators of the disturbance type samples are lower than those of HIF (high-resistance fault), and the indicators of the IC samples (inrush current samples) are the smallest. By Figure 7As can be seen from (d) in [reference], the high attribution value regions of inrush current samples are relatively evenly distributed on the entire time axis, indicating that the model's classification of such disturbance samples relies more on global features and is not limited to the waveform zero-crossing region.

[0109] Overall, Figure 7 It also shows that the indicators ZAM and KAR have a strong positive correlation, that is, when the proportion of attribution values in the waveform zero-crossing region is relatively high, the high-contribution attribution points output by the classification results tend to be more concentrated in the high region. Therefore, the above quantitative indicators jointly reflect the degree of attention of the model to key regions during classification decision-making, can be used as important quantitative indicators to measure the rationality of the model's decision-making basis, and provide more intuitive and reliable decision-making support for maintenance personnel.

[0110] Usually, the selection of model hyperparameters (such as convolution kernel size, dilation coefficient, etc.) is blind due to the lack of clear theoretical guidance. Most researchers rely on continuous combination and trial-and-error to determine appropriate hyperparameters. This process requires a large amount of computing resources and time costs, and it is difficult to obtain the optimal parameter combination. The interpretability method adopted in the present invention can intuitively analyze the internal relationship between hyperparameters and model performance, effectively avoiding blind trial-and-error while providing a clear and traceable basis for the reasonable selection of hyperparameters, and further improving the model optimization efficiency.

[0111] During the parameter optimization process of the convolution kernel size, it is found that there is a certain proportion of cases where models with different convolution kernel sizes misclassify high-resistance faults as disturbances. For a certain high-resistance fault sample, models with convolution kernel sizes of 1 and 5 misclassify it, while the model with a convolution kernel size of 3 gives the correct classification result. Figure 8 Shows the visual analysis results of the diagnosis results of the above three models with different convolution kernel sizes for this sample. From Figure 8 It can be seen that as the convolution kernel size gradually increases, the model can extract features within longer time series segments. If the convolution kernel is too small, only local features in the sample can be captured, which is too fragmented; while when the convolution kernel is set too large, it is easy to capture interference features or cause the feature region concerned by the model to shift. When the convolution kernel size is set to 3, the model can accurately cover the zero-crossing distortion region that determines the classification of the sample as a high-resistance fault, and then give the correct classification result.

[0112] It should be noted that the content not elaborated in detail in the present invention is prior art and is well-known to those skilled in the art.

[0113] Therefore, the present invention adopts the above-mentioned high-resistance fault detection and interpretability analysis method for distribution networks, which has high detection accuracy and interpretability under different operating conditions, and is suitable for high-resistance fault detection and analysis in complex distribution network environments.

[0114] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions of the present invention or make equivalent replacements, and these modifications or equivalent replacements do not enable the modified technical solutions to depart from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for detecting high-resistance faults in a distribution network and performing interpretable analysis, characterized in that It includes the following steps: Step S1: Decompose and reconstruct the original transient zero-sequence current signal by using improved adaptive noise complete ensemble empirical mode decomposition, and generate the reconstructed transient zero-sequence current signal; Step S2: Based on the reconstructed transient zero-sequence current signal generated in Step S1, construct a temporal convolutional network model and output the fault detection result; Step S3: Based on the fault detection result output in Step S2, use the fractional weighted class activation mapping method to generate an attribution heat map, analyze the correlation between the reconstructed transient zero-sequence current signal and the fault detection result, and construct a quantitative evaluation index; Step S31: The class activation mapping method generates a class activation mapping diagram by linearly weighted fusion of the feature fusion weight and the feature map; Step S32: Calculate the attribution values of each sampling point in the time series of the reconstructed transient zero-sequence current signal; Step S33: Further construct a quantitative evaluation index based on the attribution values generated in Step S32.

2. The high-resistance fault detection and interpretability analysis method for a distribution network according to claim 1, wherein Step S1 specifically includes: Step S11: Add Gaussian white noise to the original transient zero-sequence current signal, and calculate the residual value and the modal component of the first decomposition; Step S12: Repeat the step of superimposing Gaussian white noise until no further decomposition is possible, and decompose the original transient zero-sequence current signal into the sum of multiple intrinsic mode function IMF components and a residual; Select the components above IMF5 for signal superposition to form the reconstructed transient zero-sequence current signal.

3. The high-resistance fault detection and interpretability analysis method for a distribution network according to claim 1, characterized in that The temporal convolutional network model includes an input layer, multiple TCN modules, a 1×1 convolutional layer, a Flatten layer, a Dense layer, and a Softmax layer connected in sequence.

4. The high-resistance fault detection and interpretability analysis method for a distribution network according to claim 1, wherein Each TCN module includes: a causal dilated convolutional layer, a weight normalization layer, a ReLU activation function, and a Dropout unit.

5. The high-resistance fault detection and interpretability analysis method for a distribution network according to claim 1, wherein The calculation method of the class activation mapping diagram is as follows: Among them, L represents the class activation mapping graph; ReLU represents the activation function; represents the channel weight; k represents the index of the channel; c represents the category of interest; represents the activation output of the k-th channel of the l-th convolutional layer in the model.

6. The high-resistance fault detection and interpretability analysis method for a distribution network according to claim 1, wherein The steps for constructing the quantitative evaluation index are as follows: For the zero-sequence current waveform with T sampling points, define the signal zero-crossing moment as t0, and take the time series interval [t0 - ΔT, t0 + ΔT] of ΔT sampling points before and after this moment as the analysis window, and define it as the waveform zero-crossing key region set Ω1: Ω1 = {t|t ∈ [t0 - ΔT, t0 + ΔT]}; Calculate the attribution value S(t) ∈ [0, 1] of the feature at the sampling point moment t in the transient zero-sequence current time series, and define the high-attribution region set Ω2: Ω2 = {t|S(t) > τ}; where τ represents the attribution value threshold; Define the high-attribution region and key region matching degree index ZAM and the key region attribution global proportion index KAR: where |·| represents the number of elements contained in the set.

Images