Substation fault positioning analysis system based on BP neural network

Abstract

The invention relates to the technical field of fault positioning analysis, and provides a transformer substation fault positioning analysis system based on a BP neural network. The data acquisition module acquires real-time information and historical fault information of a transformer substation, the topological feature extraction module extracts topological features of the transformer substation and parameterizes the topological features, and then the feature analysis and fusion module fuses the topological features through a weighted average weight distribution method to form a comprehensive feature vector. The BP neural network module takes the comprehensive feature vector as input and constructs a multilayer neural network with a specific structure and an activation function to output fault positioning probability distribution, and the fault positioning module comprehensively determines a fault position based on the fault positioning probability distribution in combination with a voting mechanism of a K-nearest neighbor algorithm. According to the invention, through extraction and data fusion of topological features of the transformer substation, the topological features are input into the BP neural network with a specific structure to obtain fault positioning probability distribution, and then accurate fault positioning is realized in combination with a voting mechanism of a K-nearest neighbor algorithm.

Description

Technical Field

[0001] The present invention relates to the technical field of fault location analysis, and more specifically, to a substation fault location analysis system based on a BP neural network. Background Art

[0002] With the continuous expansion of the scale of the power system and the gradual increase of the voltage level, the substation, as a key hub for power transmission and distribution, its safe and stable operation is of crucial importance. Once a fault occurs in the substation, quickly and accurately locating the fault location is of great significance for reducing the power outage time, reducing economic losses, and ensuring the reliable operation of the power system.

[0003] Traditional substation fault location methods, such as the logical reasoning method based on the action information of protection and circuit breakers, gradually expose limitations when facing the increasingly complex substation network structure and diverse fault types, such as poor fault tolerance and insufficient ability to diagnose complex faults.

[0004] With the development of artificial intelligence technology, the BP neural network has been widely used in the field of substation fault location due to its powerful non-linear mapping ability and self-learning ability. It can establish a complex mapping relationship between fault characteristics and fault locations by learning a large amount of historical fault data, thereby realizing intelligent diagnosis of fault locations. At present, there have been many research results on substation fault location based on the BP neural network and have been applied to a certain extent in some substations. However, in the actual application process, the substation fault location analysis system based on the BP neural network still faces many challenges.

[0005] 1. The electrical system of the substation has a complex network topology structure, including the interconnections between various devices such as busbars, circuit breakers, disconnecting switches, and transmission lines. When a fault occurs, the fault signal will propagate and reflect in this complex network, which increases the difficulty of fault location.

[0006] 2. At the same time, the BP neural network needs to process a large number of signals from different locations and devices and accurately analyze the relationship between these signals and the fault location. However, the complex network topology makes the signal propagation path and fault characteristics become fuzzy, and the neural network may not be able to effectively extract accurate fault location information from these complex signals, resulting in inaccurate fault location. Summary of the Invention

[0007] To solve the above technical problems, the present invention provides a substation fault location analysis system based on a BP neural network. By extracting the topological features of the substation and fusing the features of the substation data, the fault location probability distribution is obtained by inputting into a BP neural network with a specific structure, and then the accurate fault location is carried out by combining the voting mechanism of the K-nearest neighbor algorithm to solve the problems in the prior art.

[0008] A substation fault location analysis system based on a BP neural network, comprising:

[0009] A data acquisition module, which deploys voltage and current sensors in each area of the substation to collect three-phase voltage and three-phase current data, and deploys temperature and gas sensors to collect temperature and gas change rate data in the substation. At the same time, based on the historical fault data of the substation, the historical fault location data of the substation, as well as the corresponding voltage and current data and temperature and gas data are collected;

[0010] A topological feature extraction module, which constructs a detailed topological map of the substation based on the design blueprint and equipment connection information of the substation, extracts topological features and parameterizes them;

[0011] A feature analysis and fusion module, which is data-connected to the data acquisition module and the topological feature extraction module. First, it extracts features from the collected voltage, current, temperature, pressure and gas data, and then fuses them with the topological feature data by a method of weighted average to allocate weights to form a comprehensive feature vector

[0012] A BP neural network module, which is data-connected to the feature analysis and fusion module. Based on the comprehensive feature vector of the feature analysis and fusion module as input, by constructing a multi-layer neural network with a specific structure and activation function, it realizes the accurate prediction of the substation fault location. At the same time, during the training process, a weighted cross-entropy loss function and a stochastic gradient descent algorithm with momentum are adopted to ensure that the network will not overfit when dealing with the problem of unbalanced sample distribution;

[0013] A fault location module, which is data-connected to the feature analysis and fusion module and the BP neural network module. Based on the fault location probability distribution calculated by the BP neural network module and the voting mechanism of the K-nearest neighbor algorithm, it conducts fault location and decision-making. It preliminarily locates according to the fault location probability distribution output by the BP neural network, uses the K-nearest neighbor algorithm to assist in location and vote, calculates a comprehensive score by combining the probability and voting results, and determines the fault location according to the comprehensive score ranking;

[0014] An interaction module, which is data-connected to the fault location module, is used to provide a visual interface and interaction functions for the system, display the substation topology diagram and fault information in 3D graphics, show the feature space distribution, and allow operation and maintenance personnel to input expert judgments and adjust system parameters.

[0015] Preferably, the topology feature extraction module includes a topology diagram construction unit and a parameterization unit;

[0016] The topology diagram construction unit constructs a topology diagram according to the design blueprint of the substation and the equipment connection information, which contains various equipment nodes and the lines connecting these equipment; assigns electrical parameters to each edge, including resistance, reactance, capacitance, admittance, and length; at the same time, for any node, adds equipment characteristic parameters according to the equipment type, including busbars, transformers, circuit breakers, and disconnectors.

[0017] The parameterization unit calculates the topology features of the nodes according to the topology diagram constructed by the topology diagram construction unit, including node degree, weighted degree centrality of the node, closeness centrality of the node, and clustering coefficient of the node.

[0018] Preferably, the feature analysis and fusion module includes an electrical feature analysis unit and a feature fusion unit;

[0019] The electrical feature analysis unit analyzes the collected three-phase voltage and three-phase current, calculates their positive-sequence, negative-sequence, zero-sequence components, as well as the total harmonic distortion rate, power factor, and asymmetry of the voltage and current;

[0020] For temperature data and pressure data, calculate the change rate of temperature and pressure. For gas data, analyze the change of gas components, and use DGA technology to analyze the change rate of the dissolved gas content in oil-immersed transformers, including hydrogen, methane, and ethane. Determine whether there is an insulation fault according to the change of different gas components.

[0021] The feature fusion unit uses the weighted average method to fuse the data analyzed and calculated by the topology features and the electrical feature analysis unit into a comprehensive feature vector according to the importance of the features.

[0022] Preferably, the BP neural network module includes a model construction unit and a model training unit;

[0023] The model construction unit is used to build the BP neural network structure. The input layer receives the comprehensive feature vector of the feature analysis and fusion module and normalizes it. Three hidden layers in a diamond structure are set, and different activation functions are used in each layer to mine features. The number of neurons in the output layer corresponds to the number of historical fault locations. After linear transformation and softmax function, the output is converted into the probability distribution of the fault location, providing a basis for fault location.

[0024] The model training unit is used to train the constructed BP neural network. For the problem of unbalanced sample distribution, a weighted cross-entropy loss function is adopted, and the random gradient descent algorithm with momentum is used to update the parameters to accelerate convergence. After dividing the data set, the training set is trained using mini-batch gradient descent. After each round of training, the performance is evaluated using the validation set, and the accuracy is calculated. An early stopping mechanism is introduced to prevent overfitting, and the learning rate is adjusted according to the performance of the validation set to improve the generalization ability and positioning accuracy of the network.

[0025] Preferably, the specific construction process of the model construction unit is as follows:

[0026] Input layer: Receive the comprehensive feature vector from the feature analysis and fusion module as input;

[0027] Let The dimension of be d, and the input layer has d neurons. First, perform preprocessing on for range normalization, and normalize each element of the feature vector to the interval [0, 1];

[0028] Hidden layer, construct three hidden layers. The first layer has n 1 neurons, the second layer has n 2 neurons, and the third layer has n 3 neurons, and n 1 < n 2 > n 3 , forming a network structure similar to a diamond shape;

[0029] For the first hidden layer, the input is x 0 = F, and the output h 1 of its first hidden layer is:

[0030] h 1 = f 1 (W 1 x 0 + b 1 )

[0031] where, W 1 is a weight matrix of d×n 1 , b 1 is a bias vector of length n 1 , and f 1 is the activation function of the first hidden layer, and the ELU activation function is adopted. Its formula is:

[0032]

[0033] For the second hidden layer, the input is x 1 = h 1, the output h of its second hidden layer 2 is:

[0034] h 2 = f 2 (W 2 x 1 + b 2 )

[0035] where W 2 is an n 1 × n 2 weight matrix, b 2 is a bias vector of length n 2 and f 2 is the activation function of the second hidden layer, using the smooth non - linear function hyperbolic tangent function, and its formula is:

[0036]

[0037] For the third hidden layer, the input is x 2 = h 2 , and the output h of its third hidden layer 2 is:

[0038] h 3 = f 3 (W 3 x 2 + b 3 )

[0039] where W 3 is an n 2 × n 3 weight matrix, b 3 is a bias vector of length n 3 and f 3 is the activation function of the third hidden layer, using the Swish function, and its formula is:

[0040] f 3 (x) = x × σ(x)

[0041] where σ(x) is the Sigmoid function, that is:

[0042]

[0043] Output layer, the number of neurons k in the output layer corresponds to the number of fault locations that occurred in the substation in the historical data. The input of the output layer is x 3 = h 3 , and the original output z of the output layer is obtained through a linear transformation:

[0044] z = W 4 x 3 + b4

[0045] Among them, W 4 is an n 3 × k weight matrix, b 4 is a bias vector of length k. To convert the original output z into a probability distribution, the softmax function is used, and the final output is:

[0046]

[0047] Among them, i and j respectively represent different fault positions, z i corresponds to the original output value of the i-th fault position, z j corresponds to the original output value of the j-th fault position, P i represents the probability of the i-th fault position.

[0048] Preferably, the model training unit takes into account the uneven sample distribution of different fault positions and uses a weighted cross-entropy loss function to train the model;

[0049] For the prediction result P i of the fault position, the loss function is:

[0050]

[0051] Among them, m is the number of samples, y ri is the true label of the r-th sample at the i-th fault position, P ri is the predicted probability of the r-th sample at the i-th fault position, ω i is the weight of the i-th fault position, which is set by the reciprocal of the number of samples at this fault position to balance the learning weights of samples at different positions;

[0052] Optimization algorithm: Use the stochastic gradient descent algorithm with momentum to update the network parameters. For the update of the weight matrix W and the bias term b, first calculate the gradients g W and g b , and the momentum update formula is:

[0053] u W = μu W - ηg W

[0054] u b = μu b - ηg b

[0055] Among them, u W and u b are momentum terms, μ is the momentum coefficient, taking 0.9, and η is the learning rate used to control the update speed of momentum update according to specific requirements;

[0056] Then update the parameters: W′ = W + u W and b′ = b + u b , where the momentum term is used to accelerate convergence and enable the network to find the optimal solution faster during training.

[0057] Preferably, when training the model using the weighted cross - entropy loss function, the dataset also needs to be divided into a training set, a validation set, and a test set. For the training set, use mini - batch gradient descent and divide the training set into multiple batches;

[0058] For each training batch, perform forward propagation to calculate the output probability, calculate the loss according to the loss function, and then calculate the gradients g W and g b , and use the stochastic gradient descent algorithm with momentum to update the network parameters;

[0059] After each round of training is completed, use the validation set to evaluate the network performance and calculate the accuracy ACC:

[0060]

[0061] where m a is the number of samples in the validation set, and argmax(P ri ) means 1 when the predicted fault location is the same as the true location, otherwise 0;

[0062] Introduce an early - stopping mechanism. If the performance of the validation set does not improve in multiple consecutive rounds, stop training early to prevent overfitting. At the same time, adjust the learning rate according to the performance of the validation set. When the performance no longer improves, reduce the learning rate to search for a better parameter space.

[0063] Preferably, the fault location module performs fault location based on the probability distribution output by the BP neural network module;

[0064] Input the real - time collected and processed feature vectors into the trained model to obtain the probability distribution of the fault location (P 1 , P 2 , P i ,..., P j ), and preliminarily determine the possibility ranking of the fault location according to the probability size;

[0065] Use the K - nearest neighbor algorithm for auxiliary location;

[0066] Store the comprehensive feature vector H hist of historical fault samples in the feature space. For the new test sample feature vector H test , calculate its Euclidean distance o from the historical samples:

[0067]

[0068] Select c nearest neighbors. If the fault location j is among the selected c nearest neighbor samples, vote for each possible fault location. Voting score:

[0069]

[0070] where, β j represents the actual fault location category of the j-th sample among the selected c nearest neighbor samples, δ is an indicator function, which is 1 when β j = i condition is satisfied, otherwise 0;

[0071] Combining probability and voting results, for the fault location i, calculate the comprehensive score:

[0072] S i = ω p P i + ω V V i

[0073] where, ω p is the weight regarding probability, ω V is the weight regarding voting score, which can be adjusted according to system performance and experience. Finally, sort the fault locations according to the comprehensive score S i and select the location with the highest score as the fault location.

[0074] Compared with the prior art, the present invention has the following beneficial effects:

[0075] 1. The present invention constructs a detailed substation topology map through the topology feature extraction module, calculates the node topology features, and the feature analysis and fusion module fuses them with the electrical features into a comprehensive feature vector. The fault location module determines the fault location based on the probability distribution output by the BP neural network and combines the voting mechanism. It realizes the effective processing of the complex network topology and multi-source signals of the substation, accurately extracts the fault information, and solves the problem of inaccurate positioning caused by complex topology in traditional and existing methods.

[0076] 2. The present invention constructs a multi-layer neural network with a specific diamond structure and hidden layers with different activation functions through the BP neural network module, converts the input comprehensive feature vector into a fault location probability distribution. During the training process, the weighted cross-entropy loss function is used to balance the learning weights of samples at different positions, the stochastic gradient descent algorithm with momentum is used to update the network parameters to accelerate convergence, and at the same time, an early stopping mechanism is introduced to prevent overfitting, and the learning rate is adjusted according to the performance of the validation set. It realizes the effective solution of the problem of unbalanced sample distribution and the improvement of the network generalization ability, more accurately learns the complex mapping relationship between fault features and fault locations, and improves the accuracy of fault location. Brief Description of the Drawings

[0077] Figure 1 is a schematic diagram of the working process of the system of the present invention;

[0078] Figure 2 is a schematic diagram of the construction process of the BP neural network model of the present invention;

[0079] Figure 3 is a schematic diagram of the specific fault location process of the present invention. Detailed Description of the Invention

[0080] The following further describes in detail the embodiments of the present invention in conjunction with the drawings and embodiments. The following embodiments are used to illustrate the present invention, but cannot be used to limit the scope of the present invention.

[0081] The present invention provides a substation fault location analysis system based on a BP neural network, including:

[0082] A data acquisition module, which deploys voltage and current sensors in each area of the substation to collect three-phase voltage and three-phase current data, and deploys sensors including temperature and gas sensors to collect temperature and gas change rate data in the substation. At the same time, based on the historical fault data of the substation, it collects the historical fault location data of the substation, as well as the corresponding voltage and current data and temperature and gas data;

[0083] A topology feature extraction module, which constructs a detailed topology map of the substation based on the design blueprint and equipment connection information of the substation, extracts topology features and parameterizes them;

[0084] A feature analysis and fusion module, which is data-connected to the data acquisition module and the topology feature extraction module. First, it extracts features from the collected voltage, current, temperature, pressure and gas data, and then fuses them with the topology feature data by the method of weighted average to allocate weights to form a comprehensive feature vector

[0085] A BP neural network module, which is data-connected to the feature analysis and fusion module. Based on the comprehensive feature vector of the feature analysis and fusion module as input, it constructs a multi-layer neural network with a specific structure and activation function to achieve accurate prediction of the fault location of the substation. At the same time, during the training process, a weighted cross-entropy loss function and a stochastic gradient descent algorithm with momentum are adopted to ensure that the network will not overfit when dealing with the problem of uneven sample distribution;

[0086] Fault location module, which is connected to the feature analysis and fusion module and the BP neural network module for data. It performs fault location and decision-making based on the fault location probability distribution calculated by the BP neural network module and the voting mechanism of the K-nearest neighbor algorithm. It preliminarily locates according to the fault location probability distribution output by the BP neural network, uses the K-nearest neighbor algorithm to assist in location and conduct voting, calculates the comprehensive score by combining the probability and voting results, and determines the fault location according to the ranking of the comprehensive score.

[0087] Interaction module, which is connected to the fault location module for data. It is used to provide the visualization interface and interaction function of the system, display the substation topology diagram and fault information in 3D graphics, show the feature space distribution, and allow the operation and maintenance personnel to input expert judgments and adjust system parameters.

[0088] Embodiment 1:

[0089] As Figures 1 - 3 shown, in this embodiment, there is a large substation in a certain area. There are many devices in the station, and the network topology structure is complex, including multiple busbars, multiple transformers, a large number of circuit breakers and disconnectors, etc. The traditional fault location method has insufficient diagnostic ability in the face of complex faults. In order to improve the accuracy and efficiency of fault location, it is decided to build a new fault location model based on the historical fault information of the substation using a substation fault location analysis system based on the BP neural network.

[0090] The data acquisition module deploys voltage and current sensors in each area of the substation to collect three-phase voltage and three-phase current data in real time. At the same time, it deploys temperature, pressure and gas sensors to collect temperature, pressure and gas change rate data of equipment such as transformers. For example, gas sensors are installed on oil-immersed transformers to monitor the changes in the content of dissolved gases such as hydrogen, methane, and ethane.

[0091] Collect the historical fault data of the substation, including historical fault location data and corresponding electrical signal data and non-electrical signal data, to provide samples for model training.

[0092] The topology construction unit constructs a topology diagram according to the design blueprint of the substation and the equipment connection information, which includes various equipment nodes and the lines connecting these equipment; assigns electrical parameters to each edge, including resistance, reactance, capacitance, admittance, and length; at the same time, for any node, according to the equipment type, including busbars, transformers, circuit breakers, disconnectors, add equipment characteristic parameters to it;

[0093] The parameterization unit calculates the topological features of the nodes according to the topology diagram constructed by the topology construction unit, including node degree, weighted degree centrality of the node, closeness centrality of the node, and clustering coefficient of the node;

[0094] The electrical feature analysis unit analyzes the collected three-phase voltages and three-phase currents, and calculates their positive-sequence, negative-sequence, and zero-sequence components, as well as the total harmonic distortion rates, power factors, and asymmetry degrees of the voltages and currents;

[0095] For the temperature data and pressure data, calculate the change rates of temperature and pressure. For the gas data, analyze the change of gas components. For example, for the change rates of the dissolved gas contents including hydrogen, methane, and ethane in an oil-immersed transformer through the DGA technology, judge whether there is an insulation fault according to the changes of different gas components;

[0096] The feature fusion unit uses the weighted average method to fuse the data analyzed and calculated by the topology feature and electrical feature analysis units into a comprehensive feature vector according to the importance of the features

[0097] The BP neural network module includes a model construction unit and a model training unit;

[0098] The model construction unit is used to build the BP neural network structure. The input layer receives the comprehensive feature vector of the feature analysis and fusion module and normalizes it. Set three hidden layers in a diamond structure. Different activation functions are used in each layer to mine features. The number of neurons in the output layer corresponds to the number of historical fault positions. After linear transformation and the softmax function, the output is converted into a fault position probability distribution, providing a basis for fault location;

[0099] The specific construction process of the model construction unit is as follows:

[0100] Input layer: Receive the comprehensive feature vector from the feature analysis and fusion module as the input;

[0101] Let have a dimension of d, and the input layer has d neurons. First, perform preprocessing on for range normalization, and normalize each element of the feature vector to the interval [0, 1];

[0102] Hidden layer, construct three hidden layers. The first layer has n 1 neurons, the second layer has n 2 neurons, and the third layer has n 3 neurons, and n 1 < n 2 > n 3 to form a network structure similar to a diamond shape;

[0103] For the first hidden layer, the input is x 0 = F, and the output h 1 of its first hidden layer is:

[0104] h1 = f 1 (W 1 x 0 + b 1 )

[0105] where W 1 is a weight matrix of d × n 1 and b 1 is a bias vector of length n 1 and f 1 is the activation function of the first hidden layer. The ELU activation function is adopted, and its formula is:

[0106]

[0107] For the second hidden layer, the input is x 1 = h 1 , and the output h 2 of the second hidden layer is:

[0108] h 2 = f 2 (W 2 x 1 + b 2 )

[0109] where W 2 is a weight matrix of n 1 × n 2 and b 2 is a bias vector of length n 2 and f 2 is the activation function of the second hidden layer. The hyperbolic tangent function, a smooth non - linear function, is adopted, and its formula is:

[0110]

[0111] For the third hidden layer, the input is x 2 = h 2 , and the output h 2 of the third hidden layer is:

[0112] h 3 = f 3 (W 3 x 2 + b 3 )

[0113] where W 3 is a weight matrix of n 2 × n 3 and b 3 is a bias vector of length n 3 and f 3The activation function of the third hidden layer is the Swish function, and its formula is:

[0114] f 3 (x) = x × σ(x)

[0115] Among them, σ(x) is the Sigmoid function, that is:

[0116]

[0117] Output layer, the number of neurons k in the output layer corresponds to the number of fault locations that appear in the substation in the historical data, and the input of the output layer is x 3 = h 3 , and the original output z of the output layer is obtained through linear transformation:

[0118] z = W 4 x 3 + b 4

[0119] Among them, W 4 is an n 3 × k weight matrix, b 4 is a bias vector with a length of k. To convert the original output z into a probability distribution, the softmax function is used, and the final output is:

[0120]

[0121] Among them, i and j respectively represent different fault locations, z i corresponds to the original output value of the i-th fault location, z j corresponds to the original output value of the j-th fault location, P i represents the probability of the i-th fault location;

[0122] Considering the unbalanced sample distribution of different fault locations, the model training unit uses the weighted cross-entropy loss function to train the model;

[0123] For the prediction result P i of the fault location, the loss function is:

[0124]

[0125] Among them, m is the number of samples, y ri is the true label of the r-th sample at the i-th fault location, P ri is the probability predicted by the r-th sample at the i-th fault location, ω i is the weight of the i-th fault location, which is set by the reciprocal of the number of samples at this fault location to balance the learning weights of samples at different locations;

[0126] Optimization algorithm, using the stochastic gradient descent algorithm with momentum to update the network parameters. For the update of the weight matrix W and the bias term b, first calculate the gradients g W and g b . The momentum update formula is:

[0127] u W = μu W - ηg W

[0128] u b = μu b - ηg b

[0129] where u W and u b are momentum terms, μ is the momentum coefficient, taking 0.9, and η is the learning rate used to control the update speed of momentum update according to specific requirements;

[0130] Then update the parameters: W′ = W + u W and b′ = b + u b . The momentum term is used to accelerate convergence, enabling the network to find the optimal solution faster during training;

[0131] When training the model using the weighted cross - entropy loss function, the dataset also needs to be divided into a training set, a validation set, and a test set. For the training set, use mini - batch gradient descent and divide the training set into multiple batches;

[0132] For each training batch, perform forward propagation to calculate the output probability, calculate the loss according to the loss function, and then calculate the gradients g W and g b through backpropagation, and use the stochastic gradient descent algorithm with momentum to update the network parameters;

[0133] After each round of training is completed, use the validation set to evaluate the network performance and calculate the accuracy ACC:

[0134]

[0135] where m a is the number of samples in the validation set, and argmax(P ri ) means 1 when the predicted fault location is the same as the true location, otherwise 0;

[0136] Introduce an early stopping mechanism. If the performance of the validation set does not improve in multiple consecutive rounds, stop training early to prevent overfitting. At the same time, adjust the learning rate according to the performance of the validation set. When the performance no longer improves, reduce the learning rate to search for a better parameter space, and finally obtain a trained BP neural network model.

[0137] A multi-layer neural network with a specific diamond structure and hidden layers with different activation functions is constructed through a BP neural network module, which converts the input comprehensive feature vector into a probability distribution of the fault location. During the training process, a weighted cross-entropy loss function is used to balance the learning weights of samples at different positions, and the stochastic gradient descent algorithm with momentum is used to update the network parameters to accelerate convergence. At the same time, an early stopping mechanism is introduced to prevent overfitting, and the learning rate is adjusted according to the performance of the validation set.

[0138] It effectively solves the problem of unbalanced sample distribution and improves the generalization ability of the network, more accurately learns the complex mapping relationship between fault features and fault locations, and improves the accuracy of fault location.

[0139] Embodiment 2:

[0140] As Figures 1 - 3 shown, in this embodiment, during the normal operation of the substation in Embodiment 1, one day the monitoring system found that the operating parameters of some equipment in the station showed abnormal fluctuations, and it was preliminarily judged that a fault might have occurred. At this time, it is necessary to use the previously established substation fault location analysis system based on the BP neural network to quickly and accurately locate the fault location, so as to take maintenance measures in time and reduce the power outage time and economic losses.

[0141] The data acquisition module is immediately started to collect voltage, current, temperature, pressure and gas data of each area of the substation in real time. At the same time, the topology feature extraction module provides the topology feature information of the current substation, and the feature analysis and fusion module extracts and fuses the features of the real-time collected data to form a comprehensive feature vector.

[0142] The processed comprehensive feature vector in real time is input into the trained BP neural network module, and through the forward propagation calculation of the network, the probability distribution of the fault location is output. For example, the probability of a fault occurring in bus A is 80%, the probability of a fault occurring in transformer B is 15%, and the probability of a fault occurring in other positions is relatively low.

[0143] The fault location module performs fault location based on the probability distribution output by the BP neural network module, stores the comprehensive feature vectors of historical fault samples in the feature space in advance, and calculates the Euclidean distance between the current test sample feature vector and the historical samples.

[0144] The feature vector collected and processed in real time is input into the trained model to obtain the probability distribution of the fault location (P 1 ,P 2 ,P i ,...,P j ), and the possibility ranking of the fault location is preliminarily determined according to the probability size;

[0145] The K-nearest neighbor algorithm is used for auxiliary positioning;

[0146] Store the comprehensive feature vector H of historical fault samples in the feature space. For the feature vector H of a new test sample hist calculate its Euclidean distance o from the historical samples: test

[0147]

[0148] Select c nearest neighbors. If the fault location j is among the selected c nearest neighbor samples, vote for each possible fault location, and the voting score is:

[0149]

[0150] where β j represents the actual fault location category of the j-th sample among the selected c nearest neighbor samples, and δ is an indicator function that is 1 when the condition β j = i is satisfied, and 0 otherwise;

[0151] Combining the probability and the voting result, calculate the comprehensive score for the fault location i:

[0152] S i = ω p P i + ω V V i

[0153] where ω p is the weight regarding the probability, and ω V is the weight regarding the voting score, which can be adjusted according to the system performance and experience. Finally, sort the fault locations according to the comprehensive score S i and select the location with the highest score as the fault location.

[0154] For example, set the weight of the probability to 0.6 and the weight of the voting score to 0.4. The probability score of bus A is relatively high, and the voting score is also high, so its comprehensive score is the highest among all possible fault locations.

[0155] Sort according to the comprehensive score, select bus A with the highest score as the fault location, and the maintenance personnel quickly go to bus A for maintenance according to the positioning result, and finally successfully detect and repair the fault, restoring the normal operation of the substation.

[0156] Based on the probability distribution output by the BP neural network, the fault location module combines the voting mechanism to determine the fault location, realizing the effective processing of the complex network topology and multi-source signals of the substation, accurately extracting fault information, and solving the problem of inaccurate positioning caused by complex topology in traditional and existing methods.

[0157] ​Embodiments of the present invention are provided for purposes of illustration and description. Although the embodiments of the present invention have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those of ordinary skill in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention.

Claims

1. A substation fault location analysis system based on BP neural network, characterized in that: include: A data acquisition module, wherein the data acquisition module deploys voltage and current sensors in various areas of the substation to collect three-phase voltage and three-phase current data, and deploys temperature and gas sensors to collect temperature and gas change rate data in the substation, and collects historical fault location data of the substation based on historical fault data of the substation, as well as corresponding voltage and current data and temperature and gas data; A topology feature extraction module, which constructs a detailed topology map of the substation based on the design blueprint of the substation and the equipment connection information and extracts the topology features to parameterize them; The feature analysis and fusion module is connected with the data acquisition module and the topological feature extraction module. It first extracts the features of the collected voltage, current, temperature, pressure and gas data, and then fuses them with the topological feature data to form a comprehensive feature vector by weighted average weight distribution. BP neural network module, the BP neural network module is connected with the feature analysis and fusion module data, based on the comprehensive feature vector of the feature analysis and fusion module as input, by constructing a multi-layer neural network with a specific structure and activation function, to achieve accurate prediction of the fault location of the substation, and at the same time, in the training process, a weighted cross entropy loss function and a stochastic gradient descent algorithm with momentum are used to ensure that the network will not overfit when dealing with the problem of uneven sample distribution; A fault location module, wherein the fault location module is connected with the feature analysis and fusion module and the BP neural network module, performs fault location and decision-making based on the fault location probability distribution calculated by the BP neural network module and the voting mechanism of the K nearest neighbor algorithm, performs preliminary location according to the fault location probability distribution output by the BP neural network, uses the K nearest neighbor algorithm to assist in location and voting, calculates a comprehensive score based on the probability and voting results, and determines the fault location based on the comprehensive score ranking; The interactive module is data-connected with the fault location module to provide a visual interface and interactive functions for the system, display the substation topology and fault information in three-dimensional graphics, display the feature space distribution, and allow operation and maintenance personnel to input expert judgment and adjust system parameters.

2. A substation fault location analysis system based on BP neural network as claimed in claim 1, characterized in that: The topological feature extraction module includes a topological map construction unit and a parameterization unit; The topology construction unit constructs a topology according to the design blueprint of the substation and the equipment connection information, which includes various equipment nodes and the lines connecting these equipment; assigns electrical parameters to each edge, including resistance, reactance, capacitance, admittance and length; at the same time, for any node, according to the equipment type, including busbar, transformer, circuit breaker, disconnector, add equipment characteristic parameters to it; The parameterization unit calculates the topological features of the nodes according to the topological graph constructed by the topological graph construction unit, including the node degree, the weighted degree centrality of the node, the proximity centrality of the node and the clustering coefficient of the node.

3. A substation fault location analysis system based on BP neural network as claimed in claim 1, characterized in that: The feature analysis and fusion module includes an electrical feature analysis unit and a feature fusion unit; The electrical characteristic analysis unit analyzes the collected three-phase voltage and three-phase current, and calculates the positive sequence, negative sequence, zero sequence components, and the total harmonic distortion rate, power factor and asymmetry of the voltage and current; For temperature data and pressure data, calculate the rate of change of temperature and pressure. For gas data, analyze the change of gas composition. Use DGA technology to calculate the change rate of dissolved gas content including hydrogen, methane and ethane in oil-immersed transformers. Determine whether there is insulation fault based on the change of different gas components. The feature fusion unit uses a weighted average method to fuse the data analyzed and calculated by the topological feature and electrical feature analysis unit into a comprehensive feature vector according to the weight assigned according to the importance of the feature.

4. A substation fault location analysis system based on BP neural network as claimed in claim 1, characterized in that: The BP neural network module includes a model building unit and a model training unit; The model building unit is used to build a BP neural network structure. The input layer receives and normalizes the comprehensive feature vector of the feature analysis and fusion module. Three hidden layers with a diamond structure are set. Different activation functions are used in each layer to mine features. The number of neurons in the output layer corresponds to the number of historical fault locations. The output is converted into a probability distribution of the fault location through linear transformation and softmax function, providing a basis for fault location. The model training unit is used to train the constructed BP neural network. To address the problem of uneven sample distribution, a weighted cross entropy loss function is adopted. The stochastic gradient descent algorithm with momentum is used to update parameters to accelerate convergence. After the data set is divided, the training set is trained using a small batch gradient descent method. After each round of training, the validation set is used to evaluate the performance and calculate the accuracy. An early stopping mechanism is introduced to prevent overfitting, and the learning rate is adjusted according to the validation set performance to improve the network generalization ability and positioning accuracy.

5. A substation fault location analysis system based on BP neural network as claimed in claim 4, characterized in that: The specific construction process of the model construction unit is as follows: Input layer: receives the comprehensive feature vector from the feature analysis and fusion module As input; set up The dimension is d, and the input layer has d neurons. First, Perform preprocessing to perform range normalization, normalizing each element of the feature vector to the interval [0,1]; Hidden layer, construct three hidden layers, the first layer has n1 neurons, the second layer has n2 neurons, the third layer has n3 neurons, and n1 <n2> n3, forming a network structure similar to a diamond shape; For the first hidden layer, the input is x0=F, and the output h1 of the first hidden layer is: h1=f1(W1x0+b1) Among them, W1 is a d×n1 weight matrix, b1 is a bias vector of length n1, f1 is the activation function of the first hidden layer, and the ELU activation function is used, and its formula is: For the second hidden layer, the input is x1=h1, and the output h2 of the second hidden layer is: h2=f2(W2x1+b2) Among them, W2 is a weight matrix of n1×n2, b2 is a bias vector of length n2, f2 is the activation function of the second hidden layer, and a smooth nonlinear function hyperbolic tangent function is used, and its formula is: For the third hidden layer, the input is x2=h2, and the output h2 of the third hidden layer is: h3=f3(W3x2+b3) Among them, W3 is a weight matrix of n2×n3, b3 is a bias vector of length n3, and f3 is the activation function of the third hidden layer, using the Swish function, whose formula is: f3(x)=x×σ(x) Among them, σ(x) is the Sigmoid function, that is: Output layer: The number of neurons k in the output layer corresponds to the number of fault locations in the substation in the historical data. The input of the output layer is x3=h3. The original output z of the output layer is obtained by linear transformation: z=W4x3+b4 Among them, W4 is a weight matrix of n3×k, b4 is a bias vector of length k, and the softmax function is used to convert the original output z into a probability distribution. The final output is: Where i and j represent different fault locations, z i The original output value corresponding to the i-th fault location, z j The original output value corresponding to the jth fault location, P i represents the probability of the i-th fault location.

6. A substation fault location analysis system based on BP neural network as claimed in claim 5, characterized in that: The model training unit takes into account the uneven distribution of samples at different fault locations and uses a weighted cross entropy loss function to train the model; The prediction result P for the fault location i , the loss function is: Where m is the number of samples, y ri is the true label of the rth sample at the i-th fault location, P ri is the probability of the rth sample being predicted at the i-th fault location, ω i is the weight of the i-th fault location, which is set by the inverse of the number of samples at the fault location to balance the learning weights of samples at different locations; The optimization algorithm uses the stochastic gradient descent algorithm with momentum to update the network parameters. For the update of the weight matrix W and the bias term b, the gradient g is first calculated W and g b , the momentum update formula is: the W =μu W -ηg W the b =μu b -ηg b where u W and u b is the momentum term, μ is the momentum coefficient, which is 0.9, and η is the learning rate used to control the update speed of momentum update according to specific needs; Then update the parameters: W′=W+u W and b′=b+u b ,The momentum term is used to accelerate convergence, allowing the network to find the optimal solution faster during training.

7. A substation fault location analysis system based on BP neural network as claimed in claim 6, characterized in that: When the weighted cross entropy loss function is used to train the model, the data set needs to be divided into a training set, a validation set, and a test set. For the training set, a small batch gradient descent is used to divide the training set into multiple batches; For each training batch, forward propagation is performed to calculate the output probability, the loss is calculated according to the loss function, and then the gradient g is calculated by back propagation. W and g b And use the stochastic gradient descent algorithm with momentum to update the network parameters; After each round of training, the validation set is used to evaluate the network performance and calculate the accuracy ACC: Where m a is the number of samples in the validation set, argmax(P ri ) indicates that the predicted fault location is 1 when it is consistent with the actual location, otherwise it is 0; An early stopping mechanism is introduced. If the performance of the validation set does not improve in multiple consecutive rounds, training is stopped early to prevent overfitting. At the same time, the learning rate is adjusted according to the performance of the validation set. When the performance no longer improves, the learning rate is reduced to find a better parameter space.

8. A substation fault location analysis system based on BP neural network as claimed in claim 1, characterized in that: The fault location module performs fault location based on the probability distribution output by the BP neural network module; The feature vectors collected and processed in real time are input into the trained model to obtain the probability distribution of the fault location (P1, P2, P i ,...,P j ), preliminarily determine the possibility ranking of the fault location according to the probability; Use K nearest neighbor algorithm for auxiliary positioning; The comprehensive feature vector H of the historical fault samples hist Stored in the feature space, for the new test sample feature vector H test , calculate its Euclidean distance o with the historical sample: Select c nearest neighbors. If the fault location j is in the selected c nearest neighbor samples, vote for each possible fault location. The voting score is: Among them, β j It indicates the actual fault location category of the jth sample among the selected c nearest neighbor samples, δ is the indicator function, when β is satisfied j =1 if condition is i, otherwise 0; Combining the probability and voting results, for the fault location i, the comprehensive score is calculated: S i =ω p P i +oh V V i Among them, ω p is the weight of probability, ω V It is about the weight of the voting score, which can be adjusted according to system performance and experience, and finally based on the comprehensive score S i Sort the fault locations and select the location with the highest score as the fault location.

Images