Distribution network switch fault automatic processing method and system

Abstract

The invention relates to the technical field of distribution network fault processing, in particular to a distribution network switch fault automatic processing method and system. When a fault diagnosis model is adopted to realize distribution network fault diagnosis, the fault diagnosis model is a fault diagnosis model of a hybrid neural network model architecture, and comprises a time sequence feature extraction branch, a switch state and topology association branch and a multi-modal feature fusion and decision branch; wherein an auxiliary consistency loss item is added into a loss function of the fault diagnosis model, the loss item encourages the model to learn a fault scene conforming to actual operation experience, for example, the situation that an interphase short circuit is predicted to a low-probability branch line tail end interval is avoided, and the loss item integrates fault types and position information through feature embedding, so that the fault diagnosis accuracy is improved. Priori knowledge is introduced by using a key scene mapping table, and finally, the fault diagnosis model prediction is ensured to accord with actual operation experience through consistency loss calculation, so that the accuracy and reliability of fault diagnosis or prediction are improved.

Description

Technical Field

[0001] This invention relates to the field of distribution network fault handling technology, and in particular to an automatic fault handling method and system for distribution network switches. Background Technology

[0002] The distribution network (DP) is a crucial link connecting the main grid and users in the power system, and its reliability directly affects users' electricity experience. DP switches are the core equipment for fault isolation and power restoration in the distribution network. Traditional DP switch fault handling relies on manual inspection, telephone communication, and manual operation, which suffers from slow response, large isolation range, and high risk of human error. Existing automation solutions generally use deep learning models to diagnose fault types and locations. However, the loss function of traditional fault diagnosis models typically uses simple weighted cross-entropy, such as fault type loss weight α=0.7 and fault interval location loss weight β=0.3, but ignores two key characteristics: the difference in fault severity and topological location sensitivity in DP switch fault scenarios. Different fault types have different impacts on system safety; for example, phase-to-phase short circuits may cause equipment burnout and require priority for accurate diagnosis; single-phase grounding usually allows short-term operation and has high diagnostic tolerance; the social impact and recovery difficulty of faults occurring at key nodes of the main line and the end nodes of branch lines differ, leading to insufficient model learning for high-risk faults and faults in critical locations.

[0003] Therefore, there is an urgent need for an automatic fault handling method for distribution network switches with high learning capabilities. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides an automatic fault handling method and system for distribution network switches, which solves the problems existing in the prior art.

[0005] This invention provides an automatic fault handling method for distribution network switches, comprising the following steps: S1: Obtain the operation data of the distribution network; S2: Perform data preprocessing operations on the operation data of the distribution network; S3: Input the preprocessed runtime data into the fault diagnosis model to achieve fault diagnosis; The fault diagnosis model is a fault diagnosis model with a hybrid neural network model architecture. The loss function of the fault diagnosis model The formula is: ; In the formula, N is the number of fault type categories; M is the number of fault interval categories; , The true labels and model predicted probabilities for each fault type are respectively classified. , These are the true labels and model-predicted probabilities for locating the fault zone, respectively. The cross-entropy loss function; , These are dynamic weighting coefficients, representing fault severity and topology sensitivity, respectively. The auxiliary consistency loss function is used to constrain the logical association between fault type and location, where λ is the balance coefficient, c is the fault type, and k is the fault interval label. h is an auxiliary topology consistency loss term. type The predicted probability vector for fault type classification; h loc The predicted probability vector for fault location; Classify loss items by fault type; Locate the loss item in the faulty section; S4: Control the switch based on the fault diagnosis results to achieve automatic fault handling.

[0006] Preferably, in step S3, the dynamic weighting coefficient The calculation formula is: ; In the formula, η is the sensitivity coefficient, and S c The severity of the fault is scored, where N is the total number of fault types and S is the number of fault types. i Score the severity of the i-th fault.

[0007] Preferably, in step S3, the dynamic weighting coefficient The calculation formula is: ; In the formula, γ T is the sensitivity coefficient. k For topological criticality scoring, M is the total number of intervals, and T is the total number of intervals. j The topological criticality score is given for the j-th interval.

[0008] Preferably, in step S3, the method for implementing the auxiliary topology consistency loss term is as follows: Sa: The predicted probability vector of the fault type classification and the predicted probability vector of the fault interval location are used as inputs to the auxiliary topology consistency loss term. Sb: Establish the prior probability relationship between fault type and fault range; Sc: Calculate the auxiliary topology consistency loss term based on the prior probability relationship between the fault type and the fault interval.

[0009] Preferably, the Sc specifically involves: calculating the joint distribution and prior distribution of the predicted probability of fault type and the predicted probability of fault interval; The formula for calculating the joint distribution of the fault type prediction probability and the fault interval prediction probability is as follows: ; The formulas for calculating the prior knowledge distributions of the predicted probability of the fault type and the predicted probability of the fault interval are as follows: ; In the formula, R c,k Indicates the fault type c In the interval k The prior probability of occurrence; The auxiliary topology consistency loss term The calculation formula is: ; In the formula, KL divergence is an asymmetric metric used to measure the distance between the predicted probability of fault type P and the predicted probability of fault interval Q.

[0010] Preferably, in Sa, the expression for the predicted probability vector of fault type classification is: ; In the formula, Let N be the predicted probability of the i-th fault type, and N be the total number of fault types. The expression for the predicted probability vector for fault location is: ; In the formula, Let M be the location prediction probability for the i-th interval, and M be the total number of intervals.

[0011] Preferably, the fault diagnosis model includes a time-series feature extraction branch, a switch state and topology association branch, and a multimodal feature fusion and decision branch; The temporal feature extraction branch includes an input layer, a 1D-CNN layer, an LSTM layer, and an output layer; the input layer is used to input preprocessed running data; the 1D-CNN layer is a 3-layer convolutional kernel structure with a kernel size of 5; the LSTM layer is a bidirectional LSTM with 64 hidden units; the output layer is used to output temporal feature vectors. The switch state and topology association branch includes a graph input layer, a graph attention layer, and a graph output layer; the graph input layer is used to input the switch opening and closing states and the adjacency matrix representing the physical connection of the switch key; the graph attention layer is a 2-layer graph attention mechanism with 4 heads, and the graph attention layer automatically learns the contribution of key switches through attention weights; the graph output layer is used to output the topology feature vector. The multimodal feature fusion and decision branch includes a feature concatenation layer, a fully connected layer, and a feature output layer. The feature concatenation layer is used to concatenate the temporal feature vector and the topological feature vector into a comprehensive feature. The fully connected layer is a two-layer fully connected structure, with Dropout applied to each layer to prevent overfitting. The feature output layer uses the Softmax activation function to classify fault types and locate fault intervals. The fault types output by the feature output layer include phase-to-phase short circuit, single-phase grounding, phase-to-phase short circuit to ground, transient fault, and normal state.

[0012] Preferably, in step S2, the data preprocessing operation includes high-frequency noise removal, sensor zero drift compensation, and abnormal data removal.

[0013] Preferably, the high-frequency noise removal is achieved using a wavelet transform-based digital filtering method, specifically: the db4 wavelet is used to perform multi-scale decomposition on the distribution network's operating data, wherein a soft thresholding function is used for thresholding of the high-frequency detail components, and the processed detail components are reconstructed with the unprocessed approximate components to obtain the denoised operating data; The sensor zero drift compensation is achieved using the sliding window mean method, which involves setting a fixed-length sliding time window, calculating the arithmetic mean of the original signal point by point within the window, and subtracting the arithmetic mean from the original signal to obtain the zero drift compensated signal. The outlier removal method employs a combination of the 3σ criterion and logical constraints to eliminate outliers. Specifically, it calculates the mean μ and standard deviation σ of the zero-drift compensated signal; and defines the normal data range as [μ...]. [3σ,μ+3σ]; If the current data point exceeds the normal data range, it is marked as an outlier. If the phase voltage of a branch is normal but the corresponding phase current suddenly changes to 0A, it is necessary to judge based on the switch position: if the switch of the branch is in the open state, then 0A is a reasonable value; if the switch is in the closed state, then 0A is a sensor fault and should be marked as abnormal; if the switch is in the closed state but all three phase currents are close to 0A and the voltage is normal, it is a false alarm or poor contact of the switch, and secondary verification needs to be triggered; in a low current grounding system, the zero-sequence current should be close to 0A during normal operation. If the zero-sequence current suddenly changes to more than 50A, it should be marked as an outlier. For the outliers, linear interpolation of the nearest normal data within a sliding window is used to replace them, thereby removing the outliers.

[0014] According to another aspect of the present invention, an automatic fault handling system for distribution network switches is provided, the system employing the above-described automatic fault handling method for distribution network switches, the system comprising: The data acquisition module is used to acquire the operational data of the distribution network; The data preprocessing module is used to perform data preprocessing operations on the operating data of the distribution network; The data diagnostic module is used to input preprocessed runtime data into the fault diagnosis model to achieve fault diagnosis. The fault handling module is used to coordinate the control of the switch based on the fault diagnosis results, thereby realizing automatic fault handling.

[0015] The embodiments of the present invention have the following technical effects: In this invention, when using a fault diagnosis model to achieve distribution network fault diagnosis, the fault diagnosis model is a hybrid neural network model architecture, which includes a time-series feature extraction branch, a switch state and topology association branch, and a multi-modal feature fusion and decision branch. The loss function of the fault diagnosis model incorporates an auxiliary consistency loss term. This loss term encourages the model to learn fault scenarios that conform to actual operating experience, such as avoiding predicting phase-to-phase short circuits to low-probability branch end sections. This loss term integrates fault type and location information through feature embedding, introduces prior knowledge using a key scenario mapping table, and finally ensures that the fault diagnosis model's predictions conform to actual operating experience through consistency loss calculation, thereby improving the accuracy and reliability of fault diagnosis or prediction. Attached Figure Description

[0016] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 This is a flowchart of an automatic fault handling method for distribution network switches provided in an embodiment of the present invention; Figure 2 This is a flowchart of the implementation method of the auxiliary topology consistency loss term provided in the embodiments of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0019] Example 1, such as Figure 1 As shown, an automatic fault handling method for distribution network switches includes the following steps: S1: Obtain the operation data of the power distribution network; The distribution network operation data includes the three-phase current, three-phase voltage, zero-sequence current, zero-sequence voltage, switch opening and closing positions, and fault action signals of each node in the distribution network.

[0020] In this step, intelligent switches integrating current transformers (CTs), voltage transformers (PTs), and position sensors, as well as distribution terminals (DTUs / FTUs), are deployed at each node of the distribution network to acquire the network's operational data.

[0021] S2: Perform data preprocessing operations on the operation data of the distribution network; The raw data collected by smart sensors in smart switches in distribution networks is often affected by high-frequency noise and sensor zero drift. This noise is caused by high-frequency harmonics generated by power electronic devices, electromagnetic radiation, or the thermal noise of the sensors themselves, and manifests as rapid fluctuations in current / voltage signals, which may mask the true fault characteristics. Furthermore, during long-term operation, the zero-point offset of sensors such as CTs / PTs causes a fixed deviation in the collected fundamental signal, affecting the sensitivity to small-current grounding faults. Therefore, data preprocessing is required for the operating data of the distribution network.

[0022] Specifically, the preprocessing operations include high-frequency noise removal, sensor zero drift compensation, and abnormal data removal.

[0023] Among them, a wavelet transform-based digital filtering method is used to remove high-frequency noise. Specifically, the db4 wavelet is used to decompose the operation data of the distribution network into multiple scales. For the high-frequency detail components, a soft threshold function is used for thresholding. The processed detail components are reconstructed with the unprocessed approximate components to obtain the denoised operation data.

[0024] Among them, the sampling sliding window mean method realizes sensor zero drift compensation; specifically, a fixed-length sliding time window is set, the arithmetic mean of the original signal within the window is calculated point by point, and the original signal is subtracted from the arithmetic mean to obtain the zero drift compensated signal.

[0025] Among the data in the distribution network operation data, abnormal data includes two categories: numerical outliers and logically contradictory values. Therefore, statistical methods are used in conjunction with physical and logical constraints to detect and remove these two types of data.

[0026] Outlier data is removed by combining the 3σ criterion with logical constraints; specifically, the mean of the signal after zero drift compensation is calculated. μ with standard deviation σ The normal data range is defined as [μ [3σ,μ+3σ]; If the current data point is outside this range, it is marked as an outlier. If the voltage of a certain phase is normal but the current of the corresponding phase suddenly changes to 0A, it is necessary to judge based on the switch position: if the switch of that branch is in the open state, then 0A is a reasonable value; if the switch is in the "closed" state, then 0A is a sensor fault and should be marked as abnormal; if the switch is in the closed state but all three phase currents are close to 0A and the voltage is normal, it is a false alarm of the switch or poor contact, and secondary verification needs to be triggered; in a low current grounding system, the zero-sequence current should be close to 0A during normal operation. If the zero-sequence current suddenly changes to more than 50A, it should be marked as an outlier. For the outliers, linear interpolation of the nearest normal data within a sliding window is used to replace them, thereby removing the outliers.

[0027] S3: Input the preprocessed runtime data into the fault diagnosis model to achieve fault diagnosis; Specifically, S3 is: S3.1: Establish the aforementioned fault diagnosis model; The fault diagnosis model includes a time-series feature extraction branch, a switch state and topology association branch, and a multimodal feature fusion and decision branch. The temporal feature extraction branch includes an input layer, a 1D-CNN layer, an LSTM layer, and an output layer; the input layer is used to input preprocessed running data; the 1D-CNN layer is a 3-layer convolutional kernel structure with a kernel size of 5; the LSTM layer is a bidirectional LSTM with 64 hidden units; the output layer is used to output temporal feature vectors. The switch state and topology association branch includes a graph input layer, a graph attention layer, and a graph output layer; the graph input layer is used to input the switch opening and closing states and the adjacency matrix representing the physical connection of the switch key; the graph attention layer is a 2-layer graph attention mechanism with 4 heads, and the graph attention layer automatically learns the contribution of key switches through attention weights; the graph output layer is used to output the topology feature vector. The multimodal feature fusion and decision branch includes a feature concatenation layer, a fully connected layer, and a feature output layer. The feature concatenation layer is used to concatenate the temporal feature vector and the topological feature vector into a comprehensive feature. The fully connected layer is a two-layer fully connected structure, with Dropout applied to each layer to prevent overfitting. The feature output layer uses the Softmax activation function to classify fault types and locate fault intervals. The fault types output by the feature output layer include phase-to-phase short circuit, single-phase grounding, phase-to-phase short circuit to ground, transient fault, and normal state.

[0028] S3.2: Train and optimize the fault diagnosis model; The training dataset of the fault diagnosis model is based on historical fault waveform data, simulation data, and manually injected fault experimental data, and each dataset is labeled with the fault type and fault location.

[0029] Traditional fault diagnosis models typically use a simple weighted cross-entropy loss function, such as a fault type loss weight α=0.7 and a fault interval location loss weight β=0.3. However, this ignores two key characteristics of distribution network fault scenarios: the difference in fault severity and the sensitivity of topological location. Different fault types have different impacts on system safety. For example, phase-to-phase short circuits may cause equipment burnout and need to be diagnosed accurately first. Single-phase grounding usually allows for short-term operation and has high diagnostic fault tolerance. The social impact and recovery difficulty of faults occurring at key nodes of the main line and the end nodes of the branch line (such as the user-side branch switch interval) are different, resulting in insufficient learning of high-risk faults and faults in key locations by the model.

[0030] To address the aforementioned issues, this embodiment proposes a loss function that incorporates fault type classification loss, fault interval location loss, and auxiliary topology consistency loss. Its formula is: ; In the formula, N is the number of fault type categories; M is the number of fault interval categories; , The true labels and model predicted probabilities for each fault type are respectively classified. , These are the true labels and model-predicted probabilities for locating the fault zone, respectively. The cross-entropy loss function; , These are dynamic weighting coefficients, representing fault severity and topology sensitivity, respectively. The auxiliary consistency loss function is used to constrain the logical association between fault type and location, where λ is the balance coefficient, c is the fault type, and k is the fault interval label. h is an auxiliary topology consistency loss term. type The predicted probability vector for fault type classification; h loc The predicted probability vector for fault location; Classify loss items by fault type; Locate the loss item in the faulty section; Among them, dynamic weights The calculation formula is: ; In the formula, η is the sensitivity coefficient, Sc is the fault severity score, phase-to-phase short circuit (c=1): S1=1.0 represents the highest severity, which is likely to cause equipment damage and large-scale power outages; phase-to-phase short circuit to ground (c=2): S2=0.9; single-phase grounding (c=3): S3=0.3 represents that the low-current grounding system is allowed to operate for a short time; transient fault (c=4): S4=0.1 represents that the fault can usually recover on its own; normal state (c=5): S5=0.0 represents that no diagnostic weight is required.

[0031] Among them, dynamic weights The calculation formula is: ; In the formula, γ T is the sensitivity coefficient. k For topology criticality scoring, the following intervals are used: main line start-up switch interval: T1=1.0; main line middle interval: T2=0.7; main line rear interval: T3=0.6; branch line interval: T4=0.4; tie switch related interval: T5=0.5.

[0032] Wherein, in the loss formula, To assist the topology consistency loss term, the logical correlation between fault type and fault interval is constrained to enhance the fault diagnosis model's understanding of the overall consistency of the fault scenario.

[0033] Among them, such as Figure 2 As shown, the implementation method of the auxiliary topology consistency loss term is as follows: Sa: The predicted probability vector for fault type classification and the predicted probability vector for fault interval location are used as inputs to the auxiliary topology consistency loss term. The expression for the predicted probability vector of the fault type classification is as follows: ; In the formula, Let N be the predicted probability of the i-th fault type, and N be the total number of fault types. The expression for the predicted probability vector for fault location is as follows: ; In the formula, Let M be the location prediction probability for the i-th interval, where M is the total number of intervals. These two vectors, used as input for subsequent steps, integrate information about the fault type and location.

[0034] Sb: Establish the prior probability relationship between fault type and fault range; A predefined topology-fault correlation matrix R is defined, with dimensions N×M, and the elements in the matrix are R. c,k Indicates the fault type c In the interval k The prior probability of occurrence; the calculation of the prior probability is based on historical data statistics. For example, the R of a phase-to-phase short circuit in the first section of the main line. 1,1 =0.8 indicates that the probability of a phase-to-phase short circuit (fault type 1) occurring in the first section (section 1) of the main line is 80%, and the probability of a phase-to-phase short circuit occurring in the last section of the branch line is R. 1,4=0.1 indicates that the probability of a phase-to-phase short circuit (fault type 1) occurring in the end section (section 4) of the branch line is 10%.

[0035] Sc: Calculate the auxiliary topology consistency loss term based on the prior probability relationship between the fault type and the fault interval; Specifically, Sc refers to: calculating the joint distribution and prior distribution of the predicted probability of the fault type and the predicted probability of the fault interval; The joint distribution of the predicted probability of the fault type and the predicted probability of the fault interval. The calculation formula is: ; Among them, the prior knowledge distribution of the predicted probability of the fault type and the predicted probability of the fault interval. The calculation formula is: ; Auxiliary topology consistency loss term The calculation formula is: ; In the formula, KL divergence is an asymmetric measure used to measure the distance between two probability distributions P and Q.

[0036] The loss formula provided in this embodiment incorporates an auxiliary consistency loss term. This loss term encourages the model to learn fault scenarios that conform to actual operating experience. For example, it avoids predicting phase-to-phase short circuits to the low-probability branch end section. This loss term integrates fault type and location information through feature embedding, introduces prior knowledge using a key scenario mapping table, and finally ensures that the fault diagnosis model prediction conforms to actual operating experience through consistency loss calculation, thereby improving the accuracy and reliability of fault diagnosis or prediction.

[0037] S3.3: Perform online fault diagnosis based on the fault diagnosis model.

[0038] The proposed fault diagnosis method based on artificial intelligence models achieves a leap from manual rule-driven to data-intelligent-driven fault diagnosis in distribution network through multimodal data fusion and a hybrid neural network architecture.

[0039] S4: Based on the fault diagnosis results, the switch is controlled in a coordinated manner to achieve automatic fault handling.

[0040] Specifically, S5 is as follows: after determining the fault section, the switch at the beginning of the fault section is immediately tripped to cut off the fault current; the switch at the end of the fault section trips after a delay of 500ms, and if the fault is permanent, the tripped state is maintained; after successful isolation, the tie switch closes after a delay of 2s to ensure that the fault point is completely de-ionized, and the switches in the non-faulty areas are closed in sequence according to the topology order.

[0041] Example 2: The present invention also provides an automatic fault handling system for distribution network switches, which adopts an automatic fault handling method for distribution network switches provided in Example 1. The system includes: The data acquisition module is used to acquire the operating data of the distribution network; The data preprocessing module is used to perform data preprocessing operations on the operating data of the distribution network; The data diagnostic module is used to input preprocessed runtime data into the fault diagnosis model to achieve fault diagnosis. The fault handling module is used to coordinate the control of the switch based on the fault diagnosis results, thereby realizing automatic fault handling.

[0042] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. An automatic fault handling method for distribution network switches, characterized in that: Includes the following steps: S1: Obtain the operation data of the distribution network; S2: Perform data preprocessing operations on the operation data of the distribution network; S3: Input the preprocessed runtime data into the fault diagnosis model to achieve fault diagnosis; The fault diagnosis model is a fault diagnosis model with a hybrid neural network model architecture. The loss function of the fault diagnosis model The formula is: ； In the formula, N is the number of fault type categories; M is the number of fault interval categories; , The true labels and model predicted probabilities for each fault type are respectively classified. , These are the true labels and model-predicted probabilities for locating the fault zone, respectively. The cross-entropy loss function; , These are dynamic weighting coefficients, representing fault severity and topology sensitivity, respectively. The auxiliary consistency loss function is used to constrain the logical association between fault type and location, where λ is the balance coefficient, c is the fault type, and k is the fault interval label. h is an auxiliary topology consistency loss term. type The predicted probability vector for fault type classification; h loc The predicted probability vector for fault location; Classify loss items by fault type; Locate the loss item in the faulty section; S4: Control the switch based on the fault diagnosis results to achieve automatic fault handling.

2. The automatic fault handling method for distribution network switches according to claim 1, characterized in that: In S3, the dynamic weighting coefficient The calculation formula is: ； In the formula, η is the sensitivity coefficient, and S c The severity of the fault is scored, where N is the total number of fault types and S is the number of fault types. i Score the severity of the i-th fault.

3. The automatic fault handling method for distribution network switches according to claim 1 or 2, characterized in that: In S3, the dynamic weighting coefficient The calculation formula is: ； In the formula, γ T is the sensitivity coefficient. k For topological criticality scoring, M is the total number of intervals, and T is the total number of intervals. j The topological criticality score is given for the j-th interval.

4. The automatic fault handling method for distribution network switches according to claim 1, characterized in that: In step S3, the method for implementing the auxiliary topology consistency loss term is as follows: Sa: The predicted probability vector of the fault type classification and the predicted probability vector of the fault interval location are used as inputs to the auxiliary topology consistency loss term. Sb: Establish the prior probability relationship between fault type and fault range; Sc: Calculate the auxiliary topology consistency loss term based on the prior probability relationship between the fault type and the fault interval.

5. The automatic fault handling method for distribution network switches according to claim 4, characterized in that: Specifically, Sc refers to: calculating the joint distribution and prior distribution of the predicted probability of fault type and the predicted probability of fault interval; The formula for calculating the joint distribution of the fault type prediction probability and the fault interval prediction probability is as follows: ； The formulas for calculating the prior knowledge distributions of the predicted probability of the fault type and the predicted probability of the fault interval are as follows: ； In the formula, R c,k Indicates the fault type c In the interval k The prior probability of occurrence; The auxiliary topology consistency loss term The calculation formula is: ； In the formula, KL divergence is an asymmetric metric used to measure the distance between the predicted probability of fault type P and the predicted probability of fault interval Q.

6. The automatic fault handling method for distribution network switches according to claim 4, characterized in that: In Sa, the expression for the predicted probability vector of fault type classification is: ； In the formula, Let N be the predicted probability of the i-th fault type, and N be the total number of fault types. The expression for the predicted probability vector for fault location is: ； In the formula, Let M be the location prediction probability for the i-th interval, and M be the total number of intervals.

7. The automatic fault handling method for distribution network switches according to claim 1, characterized in that: The fault diagnosis model includes a time-series feature extraction branch, a switch state and topology association branch, and a multimodal feature fusion and decision branch; The temporal feature extraction branch includes an input layer, a 1D-CNN layer, an LSTM layer, and an output layer; the input layer is used to input preprocessed running data; the 1D-CNN layer is a 3-layer convolutional kernel structure with a kernel size of 5; the LSTM layer is a bidirectional LSTM with 64 hidden units; the output layer is used to output temporal feature vectors. The switch state and topology association branch includes a graph input layer, a graph attention layer, and a graph output layer; the graph input layer is used to input the switch opening and closing states and the adjacency matrix representing the physical connection of the switch key; the graph attention layer is a 2-layer graph attention mechanism with 4 heads, and the graph attention layer automatically learns the contribution of key switches through attention weights; the graph output layer is used to output the topology feature vector. The multimodal feature fusion and decision branch includes a feature splicing layer, a fully connected layer, and a feature output layer; The feature splicing layer is used to splice the temporal feature vector and the topological feature vector into a comprehensive feature; the fully connected layer is a two-layer fully connected structure, with Dropout applied to each layer to prevent overfitting; the feature output layer uses the Softmax activation function to classify fault types and locate fault intervals; the fault types output by the feature output layer include phase-to-phase short circuit, single-phase grounding, phase-to-phase short circuit to grounding, transient fault, and normal state.

8. The automatic fault handling method for distribution network switches according to claim 1, characterized in that: In S2, the data preprocessing operations include high-frequency noise removal, sensor zero drift compensation, and abnormal data removal.

9. The automatic fault handling method for distribution network switches according to claim 8, characterized in that: The high-frequency noise removal is achieved using a wavelet transform-based digital filtering method. Specifically, the db4 wavelet is used to perform multi-scale decomposition on the operation data of the distribution network. For the high-frequency detail components, a soft thresholding function is used for thresholding. The processed detail components are then reconstructed with the unprocessed approximate components to obtain the denoised operation data. The sensor zero drift compensation is achieved using the sliding window mean method, which involves setting a fixed-length sliding time window, calculating the arithmetic mean of the original signal point by point within the window, and subtracting the arithmetic mean from the original signal to obtain the zero drift compensated signal. The outlier removal method employs a combination of the 3σ criterion and logical constraints to eliminate outliers. Specifically, it calculates the mean μ and standard deviation σ of the zero-drift compensated signal; and defines the normal data range as [μ...]. [3σ,μ+3σ]; If the current data point exceeds the normal data range, it is marked as an outlier. If the phase voltage of a branch is normal but the corresponding phase current suddenly changes to 0A, it is necessary to judge based on the switch position: if the switch of the branch is in the open state, then 0A is a reasonable value; if the switch is in the closed state, then 0A is a sensor fault and should be marked as abnormal; if the switch is in the closed state but all three phase currents are close to 0A and the voltage is normal, it is a false alarm or poor contact of the switch, and secondary verification needs to be triggered; in a low current grounding system, the zero-sequence current should be close to 0A during normal operation. If the zero-sequence current suddenly changes to more than 50A, it should be marked as an outlier. For the outliers, linear interpolation of the nearest normal data within a sliding window is used to replace them, thereby removing the outliers.

10. An automatic fault handling system for distribution network switches, characterized in that, The system employs an automatic fault handling method for distribution network switches as described in any one of claims 1-9, the system comprising: The data acquisition module is used to acquire the operational data of the distribution network; The data preprocessing module is used to perform data preprocessing operations on the operating data of the distribution network; The data diagnostic module is used to input preprocessed runtime data into the fault diagnosis model to achieve fault diagnosis. The fault handling module is used to coordinate the control of the switch based on the fault diagnosis results, thereby realizing automatic fault handling.

Images