A distribution network fault identification and positioning method and device

Abstract

The present invention discloses a distribution network fault identification and positioning method and device, which relates to the field of distribution network fault identification technology; after decomposing the synchronized first current signal, GAF is applied to obtain a two-dimensional image; after the synchronized first switch quantity data is uniquely encoded to obtain a numerical feature matrix; all two-dimensional images and numerical feature matrices are substituted into a deep learning model to obtain the first fault probability of each component; after building a Bayesian network based on the first switch quantity data and the first fault probability of each component, reverse reasoning is performed to determine the second fault probability of each component; the first fault probability and the second fault probability corresponding to the component are fused using the DS evidence theory to obtain the third fault probability. The fault probability is determined from the input two-dimensional image and the numerical feature matrix through a deep learning model, which improves the accuracy of fault detection, and then the first fault probability and the second fault probability are fused to improve the reliability of the overall judgment and the efficiency of the power grid fault positioning system.

Description

Technical Field

[0001] The present invention belongs to the technical field of distribution network fault identification, and in particular relates to a distribution network fault identification and positioning method and device. Background Art

[0002] The core goal of the power system is to provide users with a continuous and stable power supply. However, when a power grid fails, if it is not identified and handled in time, it may trigger a chain reaction, leading to a larger range of power grid failures, or even a large-scale power outage. In addition, power grid failures that are not identified and handled in time may also cause equipment to operate in an abnormal state for a long time, increasing equipment losses and failure risks.

[0003] Traditional distribution network fault location methods mainly rely on simple equipment such as manual inspections and fault indicators. This method is not only time-consuming and inefficient, but also requires a huge workload for manual inspections, especially in widely distributed distribution networks. Early distribution network fault location also relied on single-point detection equipment, such as current transformers and voltage transformers, which can only monitor the parameters of the installation location. When a fault occurs, it can only provide local fault information, making it difficult to quickly and accurately locate the fault point.

[0004] In order to solve the shortcomings of traditional distribution network fault location methods, modern power grid fault identification technology has gradually been developed. However, when a DC distribution network fault occurs, the existing location method relies on the fault information of a single switch quantity or electrical quantity, which makes it difficult to achieve accurate fault location, resulting in the low efficiency of the power grid fault location system. Summary of the invention

[0005] The purpose of the present invention is to solve the problem that when a fault occurs in a DC distribution network, the existing positioning method relies on the fault information of a single switch quantity or electrical quantity, which makes it difficult to accurately locate the fault, resulting in the low efficiency of the power grid fault positioning system, and proposes a distribution network fault identification and positioning method and device.

[0006] In a first aspect of the present invention, a distribution network fault identification and location method is first proposed, the method comprising:

[0007] Acquire a current signal and switch quantity data in a target distribution network, and perform data cleaning on the current signal and the switch quantity data to obtain a first current signal and a first switch quantity data;

[0008] The first current signal and the first switch quantity data are clock synchronized by using a network time protocol, the synchronized first current signal is decomposed to obtain a plurality of IMF components, and a two-dimensional image is obtained by applying GAF to each IMF component; the synchronized first switch quantity data is one-hot encoded to obtain a numerical feature matrix;

[0009] Substituting all two-dimensional images and the numerical feature matrix into a deep learning model to obtain a first fault probability of each component in the target distribution network;

[0010] Constructing a Bayesian network according to the first switch quantity data and the first fault probability of each component, and performing reverse reasoning according to the Bayesian network to determine the second fault probability of each component in the target distribution network;

[0011] For each component, the first fault probability and the second fault probability corresponding to the component are fused through the DS evidence theory to obtain the third fault probability, and the position of the component corresponding to the maximum value of the third fault probability is obtained to obtain the target fault position.

[0012] Optionally, decomposing the synchronized first current signal to obtain multiple IMF components includes:

[0013] Step 1: Add white noise to the synchronized first current signal to obtain a first signal;

[0014] Step 2: Obtain all local maxima and local minima in the first signal, and connect all local maxima and local minima by interpolation to obtain an upper envelope and a lower envelope;

[0015] Step 3: Calculate the average values ​​of the upper envelope and the lower envelope respectively to obtain an upper envelope mean and a lower envelope mean, and obtain a target mean by subtracting the lower envelope mean from the upper envelope mean;

[0016] Step 4: Subtract the target mean value from the first signal to obtain a preliminary modal component;

[0017] Step 5: Record the preliminary modal component as the first signal in step 2, and repeat steps 2 to 4 until the first preset condition is met to obtain the target IMF component;

[0018] Step 6: Subtract the target IMF component from the first current signal to obtain a second current signal, record the second current signal as the first current signal in step 1, and repeat steps 1 to 5 above until a second preset condition is met to obtain a target IMF component set.

[0019] Optionally, substituting all the two-dimensional images and the numerical feature matrix into a deep learning model to obtain a first fault probability of each element in the target distribution network includes:

[0020] Normalizing the sizes of all two-dimensional images to obtain a target two-dimensional image set, inputting any target two-dimensional image in the target two-dimensional image set into a convolutional layer of a convolutional neural network to obtain local features, and substituting the local features into multi-layer convolution and pooling processing to obtain a one-dimensional vector;

[0021] The numerical feature matrix is ​​input into the fully connected layer for linear transformation, and nonlinear feature extraction is performed through an activation function to obtain a numerical feature vector;

[0022] The one-dimensional vector and the numerical feature vector are fused to obtain a comprehensive feature vector, and all comprehensive feature vectors are converted into probability values ​​through a Sigmoid function to obtain the first failure probability of each component.

[0023] Optionally, constructing a Bayesian network according to the first switch quantity data and the first fault probability of each component, and performing reverse reasoning according to the Bayesian network to determine the second fault probability of each component in the target distribution network includes:

[0024] A Bayesian network is constructed by taking the state of each component and the state of each switch as a node in the network, wherein the directed edges between the component state nodes and the switch quantity nodes in the Bayesian network represent the dependency relationship;

[0025] The initial failure probability of each component uses the first failure probability, and the conditional probability of the switch quantity node is calculated and filled according to the historical data and the association between the component state and the switch quantity;

[0026] Fill in the corresponding conditional probability table according to the first failure probability of each component and the conditional probability of each switch node to obtain a target probability table;

[0027] The switch quantity data at the current moment is obtained, combined with the target probability table, and the second failure probability of each component is obtained through belief propagation.

[0028] Optionally, for each component, the first failure probability and the second failure probability corresponding to the component are merged by using the DS evidence theory to obtain a third failure probability, including:

[0029] The first failure probabilities corresponding to all components are converted into BPA functions and recorded as first probability distribution, and the second failure probabilities corresponding to all components are converted into BPA functions and recorded as second probability distribution;

[0030] Substituting the first probability distribution and the second probability distribution of each component into the Dempster combination rule, the third failure probability of each component is obtained.

[0031] In a second aspect of the present invention, a distribution network fault identification and positioning device is provided, comprising:

[0032] A data cleaning module, used for acquiring a current signal and switch quantity data in a target power distribution network, and performing data cleaning on the current signal and the switch quantity data to obtain a first current signal and a first switch quantity data;

[0033] a data processing module, configured to perform clock synchronization on the first current signal and the first switch quantity data through a network time protocol, decompose the synchronized first current signal to obtain a plurality of IMF components, apply GAF to each IMF component to obtain a two-dimensional image; and perform one-hot encoding on the synchronized first switch quantity data to obtain a numerical feature matrix;

[0034] A first fault probability module, used for substituting all two-dimensional images and the numerical feature matrix into a deep learning model to obtain a first fault probability of each element in the target distribution network;

[0035] A second fault probability module is used to construct a Bayesian network according to the first switch quantity data and the first fault probability of each component, and perform reverse reasoning according to the Bayesian network to determine the second fault probability of each component in the target distribution network;

[0036] The fault probability fusion module is used to fuse the first fault probability and the second fault probability corresponding to each component through the DS evidence theory to obtain the third fault probability, and obtain the position of the component corresponding to the maximum value of the third fault probability to obtain the target fault position.

[0037] Optionally, the data processing module includes:

[0038] A white noise adding module, used for adding white noise to the synchronized first current signal to obtain a first signal;

[0039] An envelope generation module, used for obtaining all local maxima and local minima in the first signal, and connecting all local maxima and local minima respectively by interpolation to obtain an upper envelope and a lower envelope;

[0040] A target mean value calculation module, used for respectively calculating the average values ​​of the upper envelope and the lower envelope to obtain an upper envelope mean value and a lower envelope mean value, and obtaining a target mean value by subtracting the lower envelope mean value from the upper envelope mean value;

[0041] A preliminary modal component module, used for obtaining a preliminary modal component by subtracting the target mean value from the first signal;

[0042] a target IMF component calculation module, used for recording the preliminary modal component as the first signal in the envelope generation module, repeating the envelope generation module to the preliminary modal component module until a first preset condition is met to obtain a target IMF component;

[0043] A loop module is used to subtract the target IMF component from the first current signal to obtain a second current signal, record the second current signal as the first current signal in the white noise adding module, and repeat the white noise adding module to the target IMF component calculation module until a second preset condition is met to obtain a target IMF component set.

[0044] Optionally, the first fault probability module includes:

[0045] A local feature extraction module, used for normalizing the sizes of all two-dimensional images to obtain a target two-dimensional image set, for any target two-dimensional image in the target two-dimensional image set, inputting the target two-dimensional image into a convolutional layer of a convolutional neural network to obtain local features, and substituting the local features into multi-layer convolution and pooling processing to obtain a one-dimensional vector;

[0046] A nonlinear feature extraction module, used for inputting the numerical feature matrix into the fully connected layer for linear transformation, and performing nonlinear feature extraction through an activation function to obtain a numerical feature vector;

[0047] The vector fusion module is used to perform vector fusion on the one-dimensional vector and the numerical feature vector to obtain a comprehensive feature vector, and convert all comprehensive feature vectors into probability values ​​through a Sigmoid function to obtain the first failure probability of each component.

[0048] Optionally, the second fault probability module includes:

[0049] A node construction module, used to construct a Bayesian network by taking the state of each component and the state of each switch as a node in the network, wherein the directed edges between the component state nodes and the switch quantity nodes in the Bayesian network represent the dependency relationship;

[0050] A conditional probability determination module, for calculating and filling the conditional probability of the switch quantity node based on the historical data and the association between the component state and the switch quantity using the first fault probability for the initial fault probability of each component;

[0051] A target probability table determination module is used to fill in the corresponding conditional probability table according to the first failure probability of each component and the conditional probability of each switch node to obtain a target probability table;

[0052] The second fault probability calculation module is used to obtain the switch quantity data at the current moment, combine the target probability table, and obtain the second fault probability of each component through belief propagation.

[0053] Optionally, the fault probability fusion module includes:

[0054] A fault probability conversion module, used to convert the first fault probabilities corresponding to all components into a BPA function recorded as a first probability distribution, and convert the second fault probabilities corresponding to all components into a BPA function recorded as a second probability distribution;

[0055] The third fault probability determination module is used to substitute the first probability distribution and the second probability distribution of each component into the Dempster combination rule to obtain a third fault probability of each component.

[0056] Beneficial effects of the present invention:

[0057] The present invention proposes a distribution network fault identification and positioning method, which obtains current signals and switch quantity data in a target distribution network, performs data cleaning on the current signals and switch quantity data respectively to obtain a first current signal and a first switch quantity data; performs clock synchronization on the first current signal and the first switch quantity data through a network time protocol, decomposes the synchronized first current signal to obtain multiple IMF components, and applies GAF to each IMF component to obtain a two-dimensional image; performs one-hot encoding on the synchronized first switch quantity data to obtain a numerical feature matrix; substitutes all the two-dimensional images and the numerical feature matrix into a deep learning model to obtain a first fault probability of each component in the target distribution network; constructs a Bayesian network according to the first switch quantity data and the first fault probability of each component, and performs reverse reasoning according to the Bayesian network to determine the second fault probability of each component in the target distribution network; for each component, the first fault probability and the second fault probability corresponding to the component are merged through the DS evidence theory to obtain a third fault probability, and the position of the component corresponding to the maximum value of the third fault probability is obtained to obtain the target fault position. By decomposing the current signal into multiple intrinsic mode functions (IMFs) through IMF decomposition, the multi-scale characteristics of the signal can be better captured. The IMF components are converted into two-dimensional images using GAF, which enhances the ability to express complex signal characteristics. The switch data is converted into a numerical feature matrix through one-hot encoding, which retains the discrete information of the data and helps to complement the characteristics of the current signal. The fault mode is automatically learned from the input two-dimensional image and numerical feature matrix through a deep learning model, which improves the accuracy of fault detection. The causal relationship between components through Bayesian network modeling can effectively capture the dependencies in complex systems. Through reverse reasoning, the judgment can be further optimized based on the existing fault probability, and the accuracy of fault location can be enhanced. By integrating the fault probability information from different sources, the first fault probability and the second fault probability, the reliability of the overall judgment is improved, and the efficiency of the power grid fault location system is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] The present invention will be further described below in conjunction with the accompanying drawings.

[0059] Figure 1A flow chart of a method for identifying and locating a distribution network fault is provided for an embodiment of the present invention;

[0060] Figure 2 A schematic structural diagram of a distribution network fault identification and locating device is provided for an embodiment of the present invention. DETAILED DESCRIPTION

[0061] Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative work shall fall within the scope of protection of the present invention.

[0062] The embodiment of the present invention provides a method for identifying and locating a distribution network fault. Figure 1 , Figure 1 A flow chart of a distribution network fault identification and location method provided by an embodiment of the present invention. The method comprises the following steps:

[0063] S101, obtaining a current signal and switch quantity data in a target distribution network, and performing data cleaning on the current signal and the switch quantity data to obtain a first current signal and a first switch quantity data;

[0064] S102, performing clock synchronization on the first current signal and the first switch quantity data through a network time protocol, decomposing the synchronized first current signal to obtain a plurality of IMF components, applying GAF to each IMF component to obtain a two-dimensional image; performing one-hot encoding on the synchronized first switch quantity data to obtain a numerical feature matrix;

[0065] S103, substituting all two-dimensional images and numerical feature matrices into a deep learning model to obtain a first fault probability of each component in the target distribution network;

[0066] S104, constructing a Bayesian network according to the first switch quantity data and the first fault probability of each component, and performing reverse reasoning according to the Bayesian network to determine the second fault probability of each component in the target distribution network;

[0067] S105, for each component, the first fault probability and the second fault probability corresponding to the component are merged by DS evidence theory to obtain a third fault probability, and the position of the component corresponding to the maximum value of the third fault probability is obtained to obtain the target fault position.

[0068] A distribution network fault identification and positioning method provided by an embodiment of the present invention decomposes the current signal into multiple intrinsic mode functions (IMFs) through IMF decomposition, which can better capture the multi-scale characteristics of the signal, and uses GAF to convert the IMF components into a two-dimensional image, thereby enhancing the ability to express complex signal characteristics. The switch quantity data is converted into a numerical feature matrix through one-hot encoding, which retains the discrete information of the data and helps to complement the characteristics of the current signal. The fault mode is automatically learned from the input two-dimensional image and the numerical feature matrix through a deep learning model, thereby improving the accuracy of fault detection. The causal relationship between components is modeled through a Bayesian network, which can effectively capture the dependencies in a complex system. Through reverse reasoning, the judgment can be further optimized based on the existing fault probability, thereby enhancing the accuracy of fault location. Then, by fusing the fault probability information from different sources, the first fault probability and the second fault probability, the reliability of the overall judgment is improved, thereby improving the efficiency of the power grid fault location system.

[0069] In one implementation, the current signal is the current measurement data at various nodes in the distribution network, such as transformers, switchgear, loads, etc.; the switch quantity refers to the status information of the switchgear, which is usually binary, that is, there are only two states: ON and OFF. In the power system, these data are used to reflect the switching status of the circuit, such as the operating status of circuit breakers, disconnectors, load switches, etc.

[0070] In one implementation, data cleaning is performed on the current signal and the switch quantity data respectively. Specifically, a low-pass filter is applied to the current signal to remove high-frequency noise, retaining the main characteristic frequency band of the fault signal, and then the signal is denoised using a wavelet transform. Wavelet denoising can effectively remove noise while retaining important signal features. Then, a moving average or exponential smoothing method is applied to smooth the denoised current signal to reduce random fluctuations in the signal. Statistical methods such as Z-score or box plots are used to detect and eliminate outliers for the switch quantity data. The outliers here are data that exceeds a threshold, and the threshold here is determined by the technician.

[0071] In one implementation, data is cleaned to ensure the quality of input data and improve the accuracy of subsequent analysis. NTP is used for clock synchronization to ensure the temporal consistency of data from different sources, which helps to improve the reliability of analysis. The current signal is converted into a two-dimensional image to capture the time and frequency characteristics of the signal, which helps the deep learning model to better understand the signal characteristics. The switch data is converted into a numerical feature matrix so that the data can be effectively processed by the deep learning model.

[0072] In one implementation, each component in the target distribution network may be a transformer, a circuit breaker, a disconnector, a distribution line, a load device, a relay protection device, or other equipment.

[0073] In one implementation, GAF is a technique for converting time series data into images, which is commonly used in image recognition and computer vision tasks. GAF can encode the angular information of the time series into a two-dimensional image.

[0074] In one implementation, the potential rules and abnormal information in the deep learning model data are used to achieve accurate prediction of faults. Then, a Bayesian network is constructed based on the switching data and the fault probability obtained by the deep learning model. The second fault probability of each component is determined by reverse reasoning, thereby providing an intuitive fault reasoning path, which helps to deeply understand the occurrence mechanism of the fault. The first fault probability obtained by the deep learning model and the second fault probability obtained by the Bayesian network are combined and fused through the DS evidence theory to obtain a more accurate third fault probability, thereby improving the accuracy and reliability of fault location.

[0075] In one embodiment, decomposing the synchronized first current signal to obtain a plurality of IMF components includes:

[0076] Step 1: Add white noise to the synchronized first current signal to obtain a first signal;

[0077] Step 2: Obtain all local maxima and local minima in the first signal, and connect all local maxima and local minima by interpolation to obtain an upper envelope and a lower envelope;

[0078] Step 3: Calculate the average values ​​of the upper envelope and the lower envelope respectively to obtain the upper envelope mean and the lower envelope mean, and subtract the lower envelope mean from the upper envelope mean to obtain the target mean;

[0079] Step 4: Subtract the target mean value from the first signal to obtain the preliminary modal component;

[0080] Step 5: Record the preliminary modal component as the first signal in step 2, and repeat steps 2 to 4 until the first preset condition is met to obtain the target IMF component;

[0081] Step 6: Subtract the target IMF component from the first current signal to obtain a second current signal, record the second current signal as the first current signal in step 1, and repeat steps 1 to 5 above until the second preset condition is met to obtain the target IMF component set.

[0082] In one implementation, by decomposing the signal into multiple IMF components, the various components in the signal can be analyzed in more detail, and changes in different frequencies and amplitudes can be identified. This fine-grained analysis helps to more accurately understand the intrinsic characteristics of the signal; adding white noise to the signal can help identify and separate the actual signal components from the noise, which can improve the ability to identify real changes in the signal and enhance the robustness to noise; by calculating the local maximum and local minimum in the signal and using the interpolation method to establish the envelope, the local characteristics of the signal can be accurately captured, which helps to identify the oscillation mode and trend changes of the signal.

[0083] In one implementation, the mean of the upper envelope and the lower envelope is calculated to obtain the target mean, which can effectively extract the main trend of the signal, remove high-frequency noise, and help extract the main features of the signal; by continuously removing modal components from the signal, different IMF components can be gradually extracted. This iterative process can ensure that more fine-grained signal features are extracted and can better capture the non-stationary characteristics of the signal; the first preset condition and the second preset condition are determined by technical personnel. For example, the first preset condition can be to stop further extraction when the number of extracted IMF components reaches 10 or to stop extracting the IMF component when the amplitude of the IMF component is less than 0.01; the second preset condition can be when the residual of the signal is less than 0.001 or when 50 iterations are reached.

[0084] In one embodiment, substituting all two-dimensional images and numerical feature matrices into a deep learning model to obtain the first fault probability of each element in the target distribution network includes:

[0085] Normalize the size of all two-dimensional images to obtain a target two-dimensional image set. For any target two-dimensional image in the target two-dimensional image set, input the target two-dimensional image into the convolution layer of the convolutional neural network to obtain local features. Substitute the local features into multi-layer convolution and pooling processing to obtain a one-dimensional vector.

[0086] The numerical feature matrix is ​​input into the fully connected layer for linear transformation, and the nonlinear feature extraction is performed through the activation function to obtain the numerical feature vector;

[0087] The one-dimensional vector and the numerical feature vector are fused to obtain a comprehensive feature vector. All comprehensive feature vectors are converted into probability values ​​through the Sigmoid function to obtain the first failure probability of each component.

[0088] In one implementation, the two-dimensional image is processed by a convolutional neural network (CNN), which can extract complex spatial features and local patterns, so that meaningful features can be extracted from the IMF component image of the current signal, which can reflect the details and changing trends of the signal; the numerical feature matrix is ​​processed by a fully connected layer, which can perform linear transformation and nonlinear feature extraction through an activation function. This makes it possible to effectively process the classification features in the switch data.

[0089] In one implementation, the image feature vector extracted from CNN is fused with the numerical feature vector extracted from the fully connected layer, which can combine two different types of data information so that the model can comprehensively consider the information in the image data and numerical data to provide a more comprehensive fault probability prediction; the activation function can be ReLU, Tanh, Softmax, etc.

[0090] In one embodiment, a Bayesian network is constructed according to the first switch quantity data and the first fault probability of each element, and performing reverse reasoning according to the Bayesian network to determine the second fault probability of each element in the target distribution network includes:

[0091] The state of each component and the state of each switch are taken as a node in the network to construct a Bayesian network. The directed edges between the component state nodes and the switch quantity nodes in the Bayesian network represent the dependency relationship.

[0092] The initial failure probability of each component uses the first failure probability, and the conditional probability of the switch quantity node is calculated and filled according to the historical data and the association between the component state and the switch quantity;

[0093] Fill in the corresponding conditional probability table according to the first failure probability of each component and the conditional probability of each switch node to obtain a target probability table;

[0094] The switch quantity data at the current moment is obtained, combined with the target probability table, and the second failure probability of each component is obtained through belief propagation.

[0095] In one implementation, the conditional dependency between component states and switch states can be clearly expressed by constructing a Bayesian network. This modeling can accurately describe the mutual influence of various parts in the system and help capture complex causal relationships; the directed edges in the network can clearly show the causal relationship between component states and switch quantities, helping to understand and analyze the dependencies between different variables in the system.

[0096] In one implementation, by combining historical data to fill in the conditional probabilities of switch nodes, existing experience and data can be fully utilized to improve the accuracy and reliability of the model; using the first failure probability of each component as the initial condition helps to introduce actual failure data and information into the model, providing a solid foundation for subsequent reasoning; by filling in the conditional probability table, the probability distribution of each node under different conditions can be systematically calculated. This makes it possible to accurately calculate the probabilities under different states in the network, improving the accuracy of reasoning; the generated target probability table can be used for further reasoning analysis to help identify possible failure modes and risks.

[0097] In one implementation, the conditional probability of the switch quantity node is calculated and filled according to the association between the historical data and the component state and the switch quantity. Specifically, according to the state of each component (for example, fault and normal), the historical data is divided into two groups: Group 1: switch quantity data when the component is in a faulty state, Group 2: switch quantity data when the component is in a normal state, Group 1: statistics on the frequency of the switch quantity state being a certain state in the case of component failure, for example, when the component fails, the frequency of the switch quantity state "on" occurring:

[0098]

[0099] Similarly, the frequency of the switch state when the component is normal is counted. For example, when the component is normal, the frequency of the switch state "on" appears:

[0100]

[0101] Where S j =k j |X i = 0 represents the probability that the switch quantity Sj takes a specific state Kj when the element Xi is in a fault state, S j =k j |X i =1 represents the case where the element Xi is in a normal state, A is the number of times the switch quantity is in state Kj, C is the total number of faults, D is the total number of normal times, and the probability of the switch quantity Sj taking a specific state Kj. Sj represents the state of the jth switch in the distribution network. The switch quantity usually has multiple possible states (such as "on" and "off"); Kj specific switch quantity state Assume that the switch quantity has two states: on (1) and off (0), then Kj can be 1 or 0; Xi represents the state of the i-th element in the distribution network. The element can be in different states, such as "fault" 0 or "normal" 1. According to the above calculation results, fill in the conditional probability table of the switch quantity node Sj in the Bayesian network.

[0102] In one implementation, the switch quantity data at the current moment is input into the network, the observed value of the switch quantity state node is set, and the initial belief value is set for all nodes. For example, for the switch quantity node, the belief value is determined according to the target probability table; for the component state node, the belief value is determined according to the historical data, the node transmitted to the current node is recorded as the child node, the adjacent node to be transmitted by the current node is recorded as the parent node, the message transmitted by each node to its parent node is calculated, and the belief value of each node is updated based on the target probability table. Each node calculates the message received from its child node, merges these messages to update the belief of the parent node, and calculates the message transmitted by the parent node to the child node. This involves passing the state probability of the parent node to the child node, and updating the belief value of each node after each round of message transmission. The belief value represents the posterior probability of the node under the observed data, repeating the message transmission and belief update steps until the preset number of iterations, which is determined by the technician, and extracting the failure probability of each component from the final belief value, which is the second failure probability of each component.

[0103] In one implementation, historical data is a record of past fault events of each component such as transformers, switches, lines, etc., including the time, location, type, impact range and processing results of the fault, and records the opening and closing status, operation time, frequency and other information of each switch in the distribution network. These data are used to calculate the conditional probability of the switch node and establish the relationship between the switch status and the component fault.

[0104] In one implementation, the current switch data is combined with the Bayesian network, which can reflect the state changes of the system in real time, making the calculation of the fault probability more accurate and timely; by combining real-time data, the fault probability prediction can be dynamically adjusted, so that the system can respond and handle faults in a timely manner. The belief propagation algorithm can efficiently calculate the posterior probability of each node in the Bayesian network. It propagates information in the network and gradually updates the probability distribution of each node, thereby obtaining the second fault probability of each component.

[0105] In one embodiment, for each component, the first failure probability and the second failure probability corresponding to the component are merged by using the DS evidence theory to obtain a third failure probability, which includes:

[0106] The first failure probabilities corresponding to all components are converted into BPA functions and recorded as first probability distribution, and the second failure probabilities corresponding to all components are converted into BPA functions and recorded as second probability distribution;

[0107] Substituting the first probability distribution and the second probability distribution of each component into the Dempster combination rule, the third failure probability of each component is obtained.

[0108] In one implementation, by fusing the first fault probability and the second fault probability, the advantages of these two different methods can be combined. The first fault probability usually reflects the prediction result of the model, while the second fault probability is a result derived based on the system state. Fusion of the two can improve the accuracy of fault detection.

[0109] In one implementation, the DST theory can reduce the errors or biases that may be introduced by a single source of evidence by combining evidence from different sources, thereby making the final failure probability more robust and reliable.

[0110] Based on the same inventive concept, the embodiment of the present invention also provides a distribution network fault identification and positioning device. Figure 2 , Figure 2 A schematic diagram of the structure of a distribution network fault identification and positioning device provided by an embodiment of the present invention includes:

[0111] A data cleaning module, used for acquiring a current signal and switch quantity data in a target distribution network, and performing data cleaning on the current signal and the switch quantity data to obtain a first current signal and a first switch quantity data;

[0112] A data processing module is used to perform clock synchronization on the first current signal and the first switch quantity data through a network time protocol, decompose the synchronized first current signal to obtain a plurality of IMF components, apply GAF to each IMF component to obtain a two-dimensional image; and perform one-hot encoding on the synchronized first switch quantity data to obtain a numerical feature matrix;

[0113] The first fault probability module is used to substitute all two-dimensional images and numerical feature matrices into the deep learning model to obtain the first fault probability of each component in the target distribution network;

[0114] A second fault probability module is used to construct a Bayesian network according to the first switch quantity data and the first fault probability of each component, and to determine the second fault probability of each component in the target distribution network by reverse reasoning according to the Bayesian network;

[0115] The fault probability fusion module is used to fuse the first fault probability and the second fault probability corresponding to each component through the DS evidence theory to obtain the third fault probability, and obtain the position of the component corresponding to the maximum value of the third fault probability to obtain the target fault position.

[0116] A distribution network fault identification and positioning device provided by an embodiment of the present invention decomposes the current signal into multiple intrinsic mode functions (IMFs) through IMF decomposition, which can better capture the multi-scale characteristics of the signal, and uses GAF to convert the IMF components into a two-dimensional image, thereby enhancing the ability to express complex signal characteristics. The switch quantity data is converted into a numerical feature matrix through one-hot encoding, which retains the discrete information of the data and helps to complement the characteristics of the current signal. The fault mode is automatically learned from the input two-dimensional image and the numerical feature matrix through a deep learning model, thereby improving the accuracy of fault detection. The causal relationship between components is modeled through a Bayesian network, which can effectively capture the dependencies in a complex system. Through reverse reasoning, the judgment can be further optimized based on the existing fault probability, thereby enhancing the accuracy of fault location. Then, by fusing the fault probability information from different sources, the first fault probability and the second fault probability, the reliability of the overall judgment is improved, thereby improving the efficiency of the power grid fault location system.

[0117] In one embodiment, the data processing module includes:

[0118] A white noise adding module, used for adding white noise to the synchronized first current signal to obtain a first signal;

[0119] An envelope generation module, used for obtaining all local maximum values ​​and local minimum values ​​in the first signal, and connecting all local maximum values ​​and local minimum values ​​by interpolation to obtain an upper envelope and a lower envelope;

[0120] A target mean value calculation module is used to calculate the average values ​​of the upper envelope and the lower envelope respectively to obtain the upper envelope mean and the lower envelope mean, and the target mean is obtained by subtracting the lower envelope mean from the upper envelope mean;

[0121] A preliminary modal component module, used for obtaining a preliminary modal component by subtracting a target mean value from a first signal;

[0122] A target IMF component calculation module is used to record the preliminary modal component as the first signal in the envelope generation module, and repeat the envelope generation module to the preliminary modal component module until the first preset condition is met to obtain the target IMF component;

[0123] A loop module is used to subtract the target IMF component from the first current signal to obtain a second current signal, record the second current signal as the first current signal in the white noise adding module, and repeat the white noise adding module to the target IMF component calculation module until the second preset condition is met to obtain the target IMF component set.

[0124] In one embodiment, the first failure probability module includes:

[0125] A local feature extraction module is used to normalize the size of all two-dimensional images to obtain a target two-dimensional image set, and for any target two-dimensional image in the target two-dimensional image set, the target two-dimensional image is input into the convolution layer of the convolutional neural network to obtain local features, and the local features are substituted into multi-layer convolution and pooling processing to obtain a one-dimensional vector;

[0126] The nonlinear feature extraction module is used to input the numerical feature matrix into the fully connected layer for linear transformation, and to extract the nonlinear features through the activation function to obtain the numerical feature vector;

[0127] The vector fusion module is used to perform vector fusion on the one-dimensional vector and the numerical feature vector to obtain a comprehensive feature vector, and convert all the comprehensive feature vectors into probability values ​​through the Sigmoid function to obtain the first failure probability of each component.

[0128] In one embodiment, the second failure probability module includes:

[0129] A node construction module is used to construct a Bayesian network by taking the state of each component and the state of each switch as a node in the network. The directed edges between the component state nodes and the switch quantity nodes in the Bayesian network represent the dependency relationship.

[0130] A conditional probability determination module, for calculating and filling the conditional probability of the switch quantity node based on the historical data and the association between the component state and the switch quantity using the first fault probability for the initial fault probability of each component;

[0131] A target probability table determination module is used to fill in the corresponding conditional probability table according to the first failure probability of each component and the conditional probability of each switch node to obtain a target probability table;

[0132] The second fault probability calculation module is used to obtain the switch quantity data at the current moment, combine it with the target probability table, and obtain the second fault probability of each component through belief propagation.

[0133] In one embodiment, the fault probability fusion module includes:

[0134] A fault probability conversion module, used to convert the first fault probabilities corresponding to all components into a BPA function recorded as a first probability distribution, and convert the second fault probabilities corresponding to all components into a BPA function recorded as a second probability distribution;

[0135] The third fault probability determination module is used to substitute the first probability distribution and the second probability distribution of each component into the Dempster combination rule to obtain a third fault probability of each component.

[0136] The above is a detailed description of an embodiment of the present invention, but the content is only a preferred embodiment of the present invention and cannot be considered to limit the scope of implementation of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. A distribution network fault identification and location method, characterized in that: The method comprises: Acquire a current signal and switch quantity data in a target distribution network, and perform data cleaning on the current signal and the switch quantity data to obtain a first current signal and a first switch quantity data; The first current signal and the first switch quantity data are clock synchronized by using a network time protocol, the synchronized first current signal is decomposed to obtain a plurality of IMF components, and a two-dimensional image is obtained by applying GAF to each IMF component; the synchronized first switch quantity data is one-hot encoded to obtain a numerical feature matrix; Substituting all two-dimensional images and the numerical feature matrix into a deep learning model to obtain a first fault probability of each component in the target distribution network; Constructing a Bayesian network according to the first switch quantity data and the first fault probability of each component, and performing reverse reasoning according to the Bayesian network to determine the second fault probability of each component in the target distribution network; For each component, the first fault probability and the second fault probability corresponding to the component are fused through the DS evidence theory to obtain the third fault probability, and the position of the component corresponding to the maximum value of the third fault probability is obtained to obtain the target fault position.

2. A distribution network fault identification and positioning method according to claim 1, characterized in that: The synchronized first current signal is decomposed to obtain multiple IMF components including: Step 1: Add white noise to the synchronized first current signal to obtain a first signal; Step 2: Obtain all local maxima and local minima in the first signal, and connect all local maxima and local minima by interpolation to obtain an upper envelope and a lower envelope; Step 3: Calculate the average values ​​of the upper envelope and the lower envelope respectively to obtain an upper envelope mean and a lower envelope mean, and obtain a target mean by subtracting the lower envelope mean from the upper envelope mean; Step 4: Subtract the target mean value from the first signal to obtain a preliminary modal component; Step 5: Record the preliminary modal component as the first signal in step 2, and repeat steps 2 to 4 until the first preset condition is met to obtain the target IMF component; Step 6: Subtract the target IMF component from the first current signal to obtain a second current signal, record the second current signal as the first current signal in step 1, and repeat steps 1 to 5 above until a second preset condition is met to obtain a target IMF component set.

3. A distribution network fault identification and positioning method according to claim 1, characterized in that: Substituting all two-dimensional images and the numerical feature matrix into the deep learning model to obtain the first fault probability of each element in the target distribution network includes: Normalizing the sizes of all two-dimensional images to obtain a target two-dimensional image set, inputting any target two-dimensional image in the target two-dimensional image set into a convolutional layer of a convolutional neural network to obtain local features, and substituting the local features into multi-layer convolution and pooling processing to obtain a one-dimensional vector; The numerical feature matrix is ​​input into the fully connected layer for linear transformation, and nonlinear feature extraction is performed through an activation function to obtain a numerical feature vector; The one-dimensional vector and the numerical feature vector are fused to obtain a comprehensive feature vector, and all comprehensive feature vectors are converted into probability values ​​through the Sigmaid function to obtain the first failure probability of each component.

4. A distribution network fault identification and positioning method according to claim 1, characterized in that: Constructing a Bayesian network according to the first switch quantity data and the first fault probability of each component, and performing reverse reasoning according to the Bayesian network to determine the second fault probability of each component in the target distribution network includes: A Bayesian network is constructed by taking the state of each component and the state of each switch as a node in the network, wherein the directed edges between the component state nodes and the switch quantity nodes in the Bayesian network represent the dependency relationship; The initial failure probability of each component uses the first failure probability, and the conditional probability of the switch quantity node is calculated and filled according to the historical data and the association between the component state and the switch quantity; Fill in the corresponding conditional probability table according to the first failure probability of each component and the conditional probability of each switch node to obtain a target probability table; The switch quantity data at the current moment is obtained, combined with the target probability table, and the second failure probability of each component is obtained through belief propagation.

5. A distribution network fault identification and positioning method according to claim 1, characterized in that: For each component, the first fault probability and the second fault probability corresponding to the component are fused through the DS evidence theory to obtain the third fault probability, which includes: The first failure probabilities corresponding to all components are converted into BPA functions and recorded as first probability distribution, and the second failure probabilities corresponding to all components are converted into BPA functions and recorded as second probability distribution; Substituting the first probability distribution and the second probability distribution of each component into the Dempster combination rule, the third failure probability of each component is obtained.

6. A distribution network fault identification and positioning device, characterized in that: The device comprises: A data cleaning module, used for acquiring a current signal and switch quantity data in a target power distribution network, and performing data cleaning on the current signal and the switch quantity data to obtain a first current signal and a first switch quantity data; a data processing module, configured to perform clock synchronization on the first current signal and the first switch quantity data through a network time protocol, decompose the synchronized first current signal to obtain a plurality of IMF components, apply GAF to each IMF component to obtain a two-dimensional image; and perform one-hot encoding on the synchronized first switch quantity data to obtain a numerical feature matrix; A first fault probability module, used for substituting all two-dimensional images and the numerical feature matrix into a deep learning model to obtain a first fault probability of each element in the target distribution network; A second fault probability module is used to construct a Bayesian network according to the first switch quantity data and the first fault probability of each component, and perform reverse reasoning according to the Bayesian network to determine the second fault probability of each component in the target distribution network; The fault probability fusion module is used to fuse the first fault probability and the second fault probability corresponding to each component through the DS evidence theory to obtain the third fault probability, and obtain the position of the component corresponding to the maximum value of the third fault probability to obtain the target fault position.

7. A distribution network fault identification and positioning device according to claim 6, characterized in that: The data processing module comprises: A white noise adding module, used for adding white noise to the synchronized first current signal to obtain a first signal; An envelope generation module, used for obtaining all local maxima and local minima in the first signal, and connecting all local maxima and local minima respectively by interpolation to obtain an upper envelope and a lower envelope; A target mean value calculation module, used for respectively calculating the average values ​​of the upper envelope and the lower envelope to obtain an upper envelope mean value and a lower envelope mean value, and obtaining a target mean value by subtracting the lower envelope mean value from the upper envelope mean value; A preliminary modal component module, configured to obtain a preliminary modal component by subtracting the target mean value from the first signal; a target IMF component calculation module, used for recording the preliminary modal component as the first signal in the envelope generation module, repeating the envelope generation module to the preliminary modal component module until a first preset condition is met to obtain a target IMF component; A loop module is used to subtract the target IMF component from the first current signal to obtain a second current signal, record the second current signal as the first current signal in the white noise adding module, and repeat the white noise adding module to the target IMF component calculation module until a second preset condition is met to obtain a target IMF component set.

8. A distribution network fault identification and positioning device according to claim 6, characterized in that: The first fault probability module includes: A local feature extraction module, used for normalizing the sizes of all two-dimensional images to obtain a target two-dimensional image set, for any target two-dimensional image in the target two-dimensional image set, inputting the target two-dimensional image into a convolutional layer of a convolutional neural network to obtain local features, and substituting the local features into multi-layer convolution and pooling processing to obtain a one-dimensional vector; A nonlinear feature extraction module, used for inputting the numerical feature matrix into the fully connected layer for linear transformation, and performing nonlinear feature extraction through an activation function to obtain a numerical feature vector; The vector fusion module is used to perform vector fusion on the one-dimensional vector and the numerical feature vector to obtain a comprehensive feature vector, and convert all comprehensive feature vectors into probability values ​​through a Sigmoid function to obtain the first failure probability of each component.

9. A distribution network fault identification and positioning device according to claim 6, characterized in that: The second fault probability module includes: A node construction module, used to construct a Bayesian network by taking the state of each component and the state of each switch as a node in the network, wherein the directed edges between the component state nodes and the switch quantity nodes in the Bayesian network represent the dependency relationship; A conditional probability determination module, for calculating and filling the conditional probability of the switch quantity node based on the historical data and the association between the component state and the switch quantity using the first fault probability for the initial fault probability of each component; A target probability table determination module is used to fill in the corresponding conditional probability table according to the first failure probability of each component and the conditional probability of each switch node to obtain a target probability table; The second fault probability calculation module is used to obtain the switch quantity data at the current moment, combine the target probability table, and obtain the second fault probability of each component through belief propagation.

10. A distribution network fault identification and positioning device according to claim 6, characterized in that: The fault probability fusion module includes: A fault probability conversion module, used to convert the first fault probabilities corresponding to all components into a BPA function recorded as a first probability distribution, and convert the second fault probabilities corresponding to all components into a BPA function recorded as a second probability distribution; The third fault probability determination module is used to substitute the first probability distribution and the second probability distribution of each component into the Dempster combination rule to obtain a third fault probability of each component.

Images