Short-circuit current zero crossing point multi-step prediction method based on double-breakpoint circuit breaker

Abstract

The invention discloses a short-circuit current zero crossing point multi-step prediction method based on a double-breakpoint circuit breaker, and belongs to the technical field of power system phase selection breaking. The method comprises the following steps: S1, data preprocessing: carrying out normalization processing on original short-circuit current data; s2, hybrid model construction: constructing a CNN-SE-GRU hybrid model based on the current data processed in the S1 in combination with a convolutional neural network, an attention mechanism and a GRU network; s3, performing CEO algorithm optimization: performing hyper-parameter optimization on the constructed hybrid model by using a CEO algorithm to obtain a CEO-CNN-SE-GRU hybrid prediction model; s4, model training: training the hybrid prediction model to obtain a stable model; and S5, performing prediction execution: calling the trained hybrid prediction model to realize multi-step prediction of the short-circuit current, and predicting a zero crossing point of the short-circuit current. Compared with the prior art, the short-circuit current zero crossing point multi-step prediction speed and precision are improved, and meanwhile better generalization is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of phase - controlled opening and closing technology in power systems, and particularly to a multi - step prediction method for the zero - crossing point of short - circuit current based on a double - break circuit breaker, which provides theoretical support for the phase - controlled breaking of the circuit breaker. Background Art

[0002] In the operation of modern power systems, the rapid and accurate disposal of short - circuit faults is a key link to ensure power supply reliability and system safety and stability. After a short - circuit fault occurs, the diversity of fault types (such as three - phase short - circuit, two - phase short - circuit, single - phase - to - ground short - circuit, etc.), the differences in system impedance parameters, and the uncertainty of the short - circuit point position in the transmission network will cause complex non - linear changes in the short - circuit current waveform. At the same time, noise interference and harmonic pollution in the power system further exacerbate the distortion degree of the short - circuit current waveform, making the accurate and rapid prediction of the short - circuit current zero - crossing point a long - standing technical problem in the power field that needs to be overcome urgently.

[0003] In traditional short - circuit current zero - crossing point prediction methods, the weighted least mean square (WLMS) algorithm predicts the zero - crossing point by continuously adjusting the filter coefficients with the goal of minimizing the mean square value of the error signal; the improved half - wave Fourier algorithm is based on the Fourier transform principle, performs spectral analysis on the short - circuit current signal, extracts the fundamental wave component and then predicts the zero - crossing point; the recursive least square (RLS) algorithm uses the least - square criterion to update parameters recursively to achieve prediction. However, these classical algorithms generally have the problem of too long sampling time. Usually, more than 10 ms of sampling data is required to complete the prediction, and the prediction error is relatively large, generally about ±1 ms, which is difficult to meet the actual requirements of the power system for rapid response to short - circuit faults.

[0004] With the wide application of deep learning technology in the field of power system fault analysis, the long short - term memory network (LSTM) algorithm has made remarkable progress in the prediction of short - circuit current zero - crossing point due to its unique gating mechanism, and the sampling time has been significantly shortened to 3 ms. Nevertheless, its prediction error still remains at ±0.5 ms. In high - voltage and extra - high - voltage power systems with extremely high precision requirements, this error range still cannot fully meet the actual engineering requirements such as rapid fault isolation and accurate opening and closing of circuit breakers.

[0005] In view of this, the present invention proposes a multi - step prediction method for the zero - crossing point of short - circuit current based on a double - break circuit breaker, which has crucial theoretical value and practical significance for improving the efficiency of power system fault handling, reducing the risk of equipment damage, and ensuring the safe and stable operation of the power grid. Summary of the Invention

[0006] The purpose of the present invention is to provide a multi-step prediction method for the zero-crossing of short-circuit current based on a double-break circuit breaker to solve the problems mentioned in the background art; the present invention effectively improves the prediction speed and accuracy of the zero-crossing of short-circuit current, has better stability and anti-noise performance, and can adapt to short-circuit fault conditions under different working conditions.

[0007] To achieve the above-mentioned invention purpose, the present invention provides the following technical solutions:

[0008] A multi-step prediction method for the zero-crossing of short-circuit current based on a double-break circuit breaker, comprising the following steps:

[0009] S1. Data preprocessing: Normalize the original short-circuit current data, scale it to the interval [0,1], keep the change trend of the current waveform unchanged, obtain the time series of the short-circuit node, and divide the obtained data into a training set, a test set and a validation set;

[0010] S2. Hybrid model construction: Based on the processed current data in S1, combine the convolutional neural network (CNN), the attention mechanism (SE) and the GRU network to construct a CNN-SE-GRU hybrid model;

[0011] S3. CEO algorithm optimization: Use the CEO algorithm to optimize the hyperparameters of the hybrid model constructed in S2 to obtain a CEO-CNN-SE-GRU hybrid prediction model with a collaborative optimization mechanism, and improve the accuracy of the zero-crossing prediction of short-circuit current;

[0012] S4. Model training: Use the training set to train the hybrid prediction model obtained in S3, and use the test set to test the trained model to obtain a stable CEO-CNN-SE-GRU hybrid prediction model;

[0013] S5. Prediction execution: Call the trained hybrid prediction model, use the data in the validation set as the model input, realize multi-step prediction of short-circuit current, and predict the zero-crossing of short-circuit current.

[0014] Preferably, the S2 specifically includes the following contents:

[0015] Feature extraction of convolutional neural network (CNN): Convert the obtained time series of the short-circuit node into a form suitable for convolutional operation through a sequence folding layer, and construct a two-dimensional input structure of the convolutional neural network (CNN). Among them, the first convolutional layer uses a 3*1 convolutional kernel to extract 64-channel features to capture local patterns and short-term dependencies of the time series; the second convolutional layer further extracts 128-channel features to capture more advanced time features;

[0016] Attention mechanism (SE) feature enhancement: Using the attention mechanism (SE), perform global average pooling on the feature sequence output by the first convolutional layer of the convolutional neural network (CNN) to compress the spatial dimension; use two fully connected layers to construct a bottleneck design with 64-32-128 nodes, and use the ReLU activation function after the first fully connected layer to introduce non-linearity; calculate the channel importance weights through the Sigmoid activation function.

[0017] Feature weighting: Combine the convolutional neural network (CNN) with the attention mechanism (SE) to initially extract local features of the short-circuit current fault waveform and perform feature weighting.

[0018] Format conversion: Restore the weighted features to a time series structure through unfolding, and use the flattening operation to compress the multi-dimensional features into a one-dimensional vector.

[0019] GRU time series modeling: Use the GRU dual-gate mechanism for time series modeling to capture long-term dependencies in the fault waveform and output the predicted waveform.

[0020] Fitness function calculation: Calculate the root mean square error RMSE between the real waveform and the predicted waveform as the fitness function of the optimization algorithm.

[0021] Preferably, the feature weighting specifically includes the following content:

[0022] The expressions of the convolutional layer and the activation function are:

[0023]

[0024] where l represents the grid layer number; M jj represents the set of positions covered by the sliding window of the convolutional kernel on the input time series; represents the output of the i-th neuron in the l-1 layer; represents the weight matrix of the l-th layer; represents the bias term of the j-th neuron in the l-th layer; represents the convolution operation; represents the weighted input of the j-th neuron in the l-th layer; represents the output feature after being processed by the activation function; f(·) represents the activation function;

[0025] The function of the channel importance differential weighting process is expressed as follows:

[0026]

[0027] where, represents the global average pooling result of the j-th channel in the l-th layer; U l represents the vector composed of all channel descriptors; H and W represent the height and width of the feature. Denotes the eigenvalue at position (h, w) of the j-th channel in the l-th layer; Denotes the output feature of the first fully connected layer; Denotes the channel attention weight; and Denote the weight matrix and bias term of the fully connected layer respectively; Denotes the recalibrated feature sequence.

[0028] Preferably, the input of the GRU time series modeling is:

[0029]

[0030] where x t Denotes the input information at the current moment; flatten denotes the flattening layer;

[0031] The function of the modeling process is expressed as:

[0032]

[0033] h t = tanh(W h [r t ⊙h t-1 , x t +g h ) (10)

[0034] h t = z t ⊙h t-1 +(1 - z t )⊙h t (11)

[0035] where h t-1 Denotes the hidden state at the previous moment; h t Denotes the hidden state passed to the next moment; h tt Is the candidate hidden state, and the activation function used is tanh; r t , z t Are the reset gate and update gate respectively; σ denotes the Sigmoid function, which is the activation function of r t , z t ; W and g are the weight and bias matrices of each control gate of the GRU network respectively; ⊙ denotes element-wise multiplication;

[0036] After the modeling is completed, the output of the fully connected layer is calculated using the hidden state of the GRU network:

[0037] y = ω out h T + b out (12)

[0038] Among them, y represents the output of the fully connected layer; ω out , b out represent the output layer weight matrix and the bias vector; h T represents the hidden state of the last time step of the sequence.

[0039] Preferably, the S3 specifically includes the following content:

[0040] Use the CEO optimization algorithm to optimize the number of GRU hidden layer nodes, Adam learning rate, and L2 regularization coefficient parameters, so that the model remains in the optimal state. The specific implementation process is as follows:

[0041] S3.1. Mutation operation

[0042] The CEO algorithm uses the chaotic mapping in Equation (13) to provide the mutation direction for each individual, and designs a search framework for the mutation operator as shown in Equation (14):

[0043]

[0044] Among them, X t+1 , respectively represent the current individual and the mutated individual; a represents the search step size; dt represents the evolution direction generated by the mapping;

[0045] S3.2. Crossover operation

[0046] After the mutation operation generates new individuals, trial vectors are generated through the binomial crossover mechanism Among them Dim represents the optimization dimension; The generation process of is as follows:

[0047]

[0048] Among them, j = 1, 2,..., Dim; j rand represents a random integer in the interval [1, Dim]; rand j (0, 1] represents a value uniformly distributed between 0 and 1 randomly taken for each j; C r represents the crossover rate;

[0049] The generation process is the same as ;

[0050] S3.3. Selection operation: The CEO algorithm uses the greedy criterion for selection, and selects the best individual from the n trial vectors obtained after the mutation and crossover operations. The function is expressed as follows:

[0051]

[0052] In the formula, are respectively the test vectors and the optimal test vector in

[0053] Preferably, the prediction sampling window of the short - circuit current zero - crossing point described in S5 is 2 ms, including 20 short - circuit current values; the prediction window is 36 ms, including 360 short - circuit current values.

[0054] The present invention further protects a computer device, which includes a processor and a memory. At least one instruction, at least one program, a code set or an instruction set is stored in the memory, and the instruction, program, code set or instruction set is loaded and executed by the processor to implement the above - mentioned multi - step prediction method for the short - circuit current zero - crossing point based on a double - break circuit breaker.

[0055] The present invention further protects a computer - readable storage medium, which is characterized in that at least one instruction, at least one program, a code set or an instruction set is stored in the computer - readable storage medium, and the instruction, program, code set or instruction set is loaded and executed by the processor to implement the above - mentioned multi - step prediction method for the short - circuit current zero - crossing point based on a double - break circuit breaker.

[0056] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0057] (1) High prediction accuracy: The prediction error of the short - circuit current zero - crossing point is controlled within ±0.4 ms, and the mean absolute error is 0.0775 ms;

[0058] (2) Short sampling time: Only 2 ms of sampling time is required to predict the short - circuit current waveform in the subsequent 36 ms;

[0059] (3) Good stability: It shows stability under different non - periodic component amplitudes, decay constants, periodic component amplitudes and fault phases;

[0060] (4) Strong anti - noise ability: It can maintain high prediction accuracy in a 20 - 50 dB noise environment, and the maximum mean prediction error is only 0.32 ms. Description of the Drawings

[0061] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings involved in the embodiments are briefly introduced below. Obviously, the drawings described below are only schematic illustrations of some embodiments of the present invention. For those skilled in the art, other forms of drawings can also be constructed based on these drawings without creative labor.

[0062] Figure 1 It is the framework diagram of the CNN - SE - GRU model proposed in Embodiment 1 of the present invention;

[0063] Figure 2 This is the training framework diagram of the CEO-optimized CNN-SE-GRU model proposed in Embodiment 1 of the present invention;

[0064] Figure 3 This is the schematic diagram of the simulation prediction results of the CEO-CNN-SE-GRU model proposed in Embodiment 1 of the present invention;

[0065] Figure 4 This is the comparison diagram of the prediction results of multiple models proposed in Embodiment 2 of the present invention;

[0066] Figure 5 This is the statistical chart of the prediction error of the CEO-CNN-SE-GRU model proposed in Embodiment 2 of the present invention;

[0067] Figure 6 This is the analysis diagram of the error influence of the CEO-CNN-SE-GRU model proposed in Embodiment 2 of the present invention;

[0068] Figure 7 This is the analysis diagram of the noise influence of the CEO-CNN-SE-GRU model proposed in Embodiment 2 of the present invention. Detailed implementation manners

[0069] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments.

[0070] The present invention proposes a multi-step prediction method for the short-circuit current zero-crossing based on a double-break circuit breaker, which organically combines the local feature extraction advantage of the convolutional neural network (CNN), the key feature enhancement characteristic of the attention mechanism, the time series modeling ability of the GRU network, and the chaotic evolutionary optimization algorithm (CEO) to construct a CEO-CNN-SE-GRU hybrid prediction model with a collaborative optimization mechanism. Next, the multi-step prediction method for the short-circuit current zero-crossing based on the double-break circuit breaker proposed by the present invention will be described in conjunction with the relevant accompanying drawings and specific examples, and the specific content is as follows.

[0071] Embodiment 1:

[0072] The present invention proposes a multi-step prediction method for the short-circuit current zero-crossing based on a double-break circuit breaker, including:

[0073] Step 1, data preprocessing: Normalize the original short-circuit current data, scale it to the interval [0, 1], keep the change trend of the current waveform unchanged, obtain the time series of the short-circuit node, and divide the obtained data into a training set, a test set, and a validation set.

[0074] Step 2, Hybrid Model Construction: Based on the processed current data in S1, a CNN-SE-GRU hybrid model is constructed by combining a Convolutional Neural Network (CNN), a Squeeze-and-Excitation (SE) mechanism, and a GRU network; please refer to Figure 1 , which specifically includes the following content:

[0075] Feature Extraction of Convolutional Neural Network (CNN): The time series of the obtained short-circuit nodes is converted into a form suitable for convolution operations through a sequence folding layer, and a two-dimensional input structure of the Convolutional Neural Network (CNN) is constructed. Among them, the first convolutional layer uses a 3*1 convolutional kernel to extract 64-channel features, capturing local patterns and short-term dependencies of the time series; the second convolutional layer further extracts 128-channel features, capturing more advanced time features;

[0076] Feature Enhancement of Squeeze-and-Excitation (SE) Mechanism: The feature sequence output by the first convolutional layer of the Convolutional Neural Network (CNN) is globally average pooled using the Squeeze-and-Excitation (SE) mechanism to compress the spatial dimension; a bottleneck design with 64-32-128 nodes is constructed using two fully connected layers, and the ReLU activation function is used after the first fully connected layer to introduce non-linearity; the channel importance weights are calculated through the Sigmoid activation function;

[0077] Feature Weighting: The Convolutional Neural Network (CNN) is combined with the Squeeze-and-Excitation (SE) mechanism to initially extract local features of the short-circuit current fault waveform and perform feature weighting; specifically including:

[0078] The expressions of the convolutional layer and the activation function are as follows:

[0079]

[0080] where l represents the number of grid layers; M jj represents the set of positions covered by the sliding window of the convolutional kernel on the input time series; represents the output of the i-th neuron in the (l-1) layer; represents the weight matrix of the l-th layer; represents the bias term of the j-th neuron in the l-th layer; represents the convolution operation; represents the weighted input of the j-th neuron in the l-th layer; represents the output feature after being processed by the activation function; f(·) represents the activation function;

[0081] The function representation of the channel importance differential weighting process is as follows:

[0082]

[0083] where represents the global average pooling result of the j-th channel in the l-th layer; Ul Denote the vector composed of all channel descriptors; H and W denote the height and width of the feature; Denote the feature value of the j-th channel in the l-th layer at the position (h, w); Denote the output feature of the first fully connected layer; Denote the channel attention weight; and Denote the weight matrix and bias term of the fully connected layer respectively; Denote the recalibrated feature sequence;

[0084] Format conversion: Restore the weighted feature to the time series structure through unfolding, and compress the multi-dimensional feature into a one-dimensional vector by using the flattening operation;

[0085] GRU time series modeling: Use the GRU dual gating mechanism for time series modeling to capture the long-term dependencies in the fault waveform. The fully connected layer maps the output of the GRU to a 360-dimensional vector, and calculates the final 360-step prediction output through the regression layer; The input of the GRU time series modeling is:

[0086]

[0087] where, x t Denote the input information at the current moment;

[0088] The function of the modeling process is expressed as:

[0089]

[0090] h t = tanh(W h [r t ⊙h t-1 x t + g h ) (10)

[0091] h t = z t ⊙h t-1 +(1 - z t )⊙h t (11)

[0092] where, h t-1 Denote the hidden state at the previous moment; h t Denote the hidden state passed to the next moment; h tt is the candidate hidden state, and the activation function used is tanh; r t 、z t are the reset gate and update gate respectively; σ denotes the Sigmoid function, which is r t 、z tActivation function; W and g are the weight and bias matrices of each control gate of the GRU network respectively;

[0093] After modeling, use the hidden state of the GRU network to calculate the output of the fully connected layer:

[0094] y = ω out h T + b out (12)

[0095] where y represents the output of the fully connected layer; ω out , b out represent the weight matrix and bias vector of the output layer; h T represents the hidden state at the last time step of the sequence;

[0096] Fitness function calculation: Calculate the root mean square error RMSE between the real waveform and the predicted waveform as the fitness function of the optimization algorithm.

[0097] Step 3. CEO algorithm optimization: The CEO algorithm generates a chaotic sequence through the exponential discrete memristor mapping (EMD) on the structural framework of the traditional DE algorithm, showing hyperchaotic characteristics when k = 2.66. The CEO algorithm diversifies the search space, avoiding the problems of local optimum and search stagnation. The CEO algorithm optimizes the hyperparameters of the CNN-SE-GRU hybrid model, avoiding the cumbersome process of manual tuning. The fitness function selects the RMSE calculated by the CNN-SE-GRU model, and the initial chaotic search range is defined as [-0.5, -0.25] and [0.5, 0.25]. The three hyperparameters to be optimized are the learning rate of the Adam optimizer, the number of nodes in the GRU hidden layer, and the L2 regularization coefficient. Please refer to Figure 2 for the specific implementation process as follows:

[0098] First, initialize the population and randomly generate 6 groups of solutions, and calculate the fitness of the initial population by substituting them into the CNN-SE-GRU hybrid model respectively to find the current optimal hyperparameter combination.

[0099] Then, randomly pair the population individuals for iterative optimization. In each iteration, first map the parameter space to the chaotic space, then use the EMD mapping to generate a chaotic sequence, and map the chaotic sequence back to the parameter space through the inverse mapping.

[0100]

[0101] The mutation operation is carried out by combining the current mutation strategy (Equation 18) with the optimal mutation strategy (Equation 19), and then the trial vectors are generated through the binomial crossover operation. The CNN-SE-GRU hybrid model is constructed using each trial vector respectively and the fitness value is calculated. The selection operation retains the better combination of the initial fitness and the fitness of the trial vectors, and the global optimal solution is updated after each iteration. When the fitness change is less than 1×10 -8 or the maximum number of iterations is reached, the algorithm terminates. The CEO-CNN-SE-GRU hybrid model is constructed using the optimal hyperparameter combination. By combining chaos theory and evolutionary algorithm, the CEO balances the global exploration and local exploitation capabilities, effectively improving the accuracy of short-circuit current zero-crossing prediction.

[0102] Step 4, Model training: The hybrid prediction model obtained in S3 is trained using the training set, and the trained model is tested using the test set to obtain a stable CEO-CNN-SE-GRU hybrid prediction model;

[0103] Step 5, Prediction execution: The prediction sampling window is set to 2 ms, including 20 short-circuit current values; the prediction window is set to 36 ms, including 360 short-circuit current values; the trained hybrid prediction model is called, and the data in the validation set is used as the model input to achieve multi-step prediction of the short-circuit current, and predict the zero-crossing point of the short-circuit current (the results are shown in Table 1 and Figure 3 as shown).

[0104] Table 1 Zero-point and extreme value errors of the CEO-CNN-SE-GRU multi-step prediction model

[0105]

[0106] The statistical results of the zero-point and peak errors between the short-circuit current prediction waveform and the true waveform are shown in Table 1. Combining Table 1, it can be seen that the zero-point error at t3 is the largest, and the absolute error reaches 0.17 ms; the error at t2 is the smallest, only 0.01 ms; the average absolute error and average relative error at the zero-point are 0.0775 ms and 0.6075% respectively; the relative error at the extreme value P2 is 5.09%, and the error in the occurrence time of the extreme value point is within 0.1 ms.

[0107] Example 2:

[0108] Based on Example 1 but with differences, control experiments and characterization experiments are designed to characterize the performance of the multi-step prediction method for short-circuit current zero-crossing based on the double-break circuit breaker proposed in the present invention and the CEO-CNN-SE-GRU hybrid prediction model constructed, and the characterization results are as Figures 4-7 shown.

[0109] Figure 4 is a comparison chart of multi-model prediction results. According toFigure 4 It can be seen that the prediction result of the CEO-CNN-SE-GRU hybrid prediction model constructed in this example has the highest degree of fitting with the true value and the highest prediction accuracy.

[0110] Combined with Figures 5-7 It can be seen that the short-circuit current zero-crossing multi-step prediction method based on a double-break circuit breaker proposed in the present invention:

[0111] 1) High prediction accuracy: The prediction error of the short-circuit current zero-crossing is controlled within ±0.4 ms, and the mean absolute error is 0.0775 ms;

[0112] 2) Short sampling time: Only 2 ms of sampling time is required to predict the short-circuit current waveform for the subsequent 36 ms;

[0113] 3) Good stability: It shows stability under different magnitudes of non-periodic components, decay constants, magnitudes of periodic components, and fault phases;

[0114] 4) Strong anti-noise ability: It can maintain high prediction accuracy in a noise environment of 20 - 50 dB, and the maximum average prediction error is only 0.32 ms.

[0115] In order to verify the generalization performance of the model, 140 groups of short-circuit current fault waveforms were randomly generated for verification, and the overall error was controlled within 0.4 ms. By separately studying the influence of changes in factors such as fault phase angle, magnitudes of each component, decay constant of the DC component, and signal-to-noise ratio on the model prediction error, the generalization ability of the hybrid prediction model was verified.

[0116] It should be noted that in this invention patent, relative terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations.

[0117] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent substitutions or changes, should be covered by the protection scope of the present invention.

Claims

1. A multi-step prediction method for short-circuit current zero-crossing point based on a double-breakpoint circuit breaker, characterized in that: The following steps are involved: S1. Data preprocessing: normalize the original short-circuit current data, scale it to the interval [0, 1], keep the current waveform change trend unchanged, obtain the time series of the short-circuit node, and divide the obtained data into training set, test set and validation set; S2, hybrid model construction: Based on the current data processed in S1, the CNN-SE-GRU hybrid model is constructed by combining the convolutional neural network, attention mechanism and GRU network; S3, CEO algorithm optimization: Use the CEO algorithm to optimize the hyperparameters of the hybrid model constructed in S2, and obtain a CEO-CNN-SE-GRU hybrid prediction model with a collaborative optimization mechanism to improve the accuracy of short-circuit current zero-crossing point prediction; S4, model training: use the training set to train the hybrid prediction model obtained in S3, and use the test set to test the trained model to obtain a stable CEO-CNN-SE-GRU hybrid prediction model; S5. Prediction execution: Call the trained hybrid prediction model and use the data in the validation set as the model input to achieve multi-step prediction of short-circuit current and predict the zero-crossing point of short-circuit current.

2. A multi-step prediction method for short-circuit current zero-crossing point based on a double-breakpoint circuit breaker according to claim 1, characterized in that: The S2 specifically includes the following contents: Convolutional neural network feature extraction: The obtained time series of short-circuit nodes is converted into a form suitable for convolution operation through the sequence folding layer, and a two-dimensional input structure of the convolutional neural network is constructed. The first convolution layer uses a 3*1 convolution kernel to extract 64-channel features to capture the local patterns and short-term dependencies of the time series; the second convolution layer further extracts 128-channel features to capture more advanced temporal features; Attention mechanism feature enhancement: The feature sequence output by the first convolutional layer of the convolutional neural network is globally averaged and pooled using the attention mechanism to compress the spatial dimension. A bottleneck design of 64-32-128 nodes is constructed using two fully connected layers, and the ReLU activation function is used to introduce nonlinearity after the first fully connected layer. The channel importance weight is calculated using the Sigmoid activation function. Feature weighting: The convolutional neural network is combined with the attention mechanism to preliminarily extract the local features of the short-circuit current fault waveform and perform feature weighting; Format conversion: The weighted features are restored to a time series structure through unfolding, and the multi-dimensional features are compressed into a one-dimensional vector through flattening operation; GRU timing modeling: GRU dual-gating mechanism is used for timing modeling to capture the long-term dependencies in the fault waveform and output the predicted waveform; Fitness function calculation: The root mean square error (RMSE) is calculated using the real waveform and the predicted waveform as the fitness function of the optimization algorithm.

3. A multi-step prediction method for short-circuit current zero-crossing point based on a double-breakpoint circuit breaker according to claim 2, characterized in that: The feature weighting specifically includes the following contents: The expression of convolution layer and activation function is: Where, l represents the number of grid layers; M j Represents the set of positions covered by the sliding window of the convolution kernel on the input time series; Represents the output of the i-th neuron in layer l-1; represents the weight matrix of the lth layer; represents the bias term of the jth neuron in the lth layer; Represents the convolution operation; represents the weighted input of the jth neuron in layer l; represents the output feature after being processed by the activation function; f(·) represents the activation function; The function of the channel importance differential weighting process is expressed as follows: in, represents the global average pooling result of the jth channel of layer l; U l Represents the vector composed of all channel descriptors; H and W represent the height and width of the feature; Represents the eigenvalue of the jth channel of layer l at position (h, w); Represents the output features of the first fully connected layer; represents the channel attention weight; and Represent the weight matrix and bias term of the fully connected layer respectively; Represents the recalibrated feature sequence.

4. A multi-step prediction method for short-circuit current zero-crossing point based on a double-breakpoint circuit breaker according to claim 3, characterized in that: The input of the GRU timing modeling is: Among them, x t Indicates the input information at the current moment; flatten means flattening the layer; The function of the modeling process is expressed as: Among them, h t-1 Indicates the hidden state of the previous moment; h t Indicates the hidden state passed to the next moment; h tt is the candidate hidden state, and the activation function used is tanh; r t 、z t are reset gate and update gate respectively; σ represents Sigmoid function, which is r t 、z t The activation function; W and g are the weight and bias matrices of each control gate of the GRU network respectively; ⊙ represents element-by-element multiplication; After the modeling is completed, the GRU network hidden state is used to calculate the fully connected layer output: y=ω out h T +b out (12) Where y represents the output of the fully connected layer; ω out 、b out represents the output layer weight matrix and bias vector; h T Represents the hidden state at the last time step of the sequence.

5. The short-circuit current zero-crossing multi-step prediction method based on a double-breakpoint circuit breaker according to claim 4 is characterized in that: The S3 specifically includes the following contents: The CEO optimization algorithm is used to optimize the number of GRU hidden layer nodes, Adam learning rate and L2 regularization coefficient parameters to keep the model in the optimal state. The specific implementation process is as follows: S3.

1. Mutation Operation The CEO algorithm uses the chaotic mapping of formula (13) to provide a mutation direction for each individual and designs a search framework of the mutation operator as shown in formula (14): Among them, X t+1 , They represent the current individual and the individual after mutation respectively; a represents the search step length; dt represents the evolutionary direction generated by the mapping; S3.2 Crossover Operation After the mutation operation generates new individuals, the test vector is generated through the binomial crossover mechanism in Dim represents the optimization dimension; The generation process is as follows: Where, j = 1, 2, ..., Dim; j rand Represents a random integer in the interval [1, Dim]; rand j (0, 1] means that each j randomly takes a uniformly distributed value between 0 and 1; C r represents the crossover rate; The generation process and Consistency; S3.3, Selection operation: The CEO algorithm uses the greedy criterion to select the best individual from the n test vectors obtained after mutation and crossover operations. The function is expressed as follows: In the formula, The test vectors are The optimal trial vector in .

6. The short-circuit current zero-crossing point multi-step prediction method based on a double-breakpoint circuit breaker according to claim 5, characterized in that: The prediction sampling window of the short-circuit current zero-crossing point in S5 is 2ms, including 20 short-circuit current values; the prediction window is 36ms, including 360 short-circuit current values.

7. A computer device, characterized in that: The computer device includes a processor and a memory, wherein the memory stores at least one instruction, at least one program, code set or instruction set, and the instruction, program, code set or instruction set is loaded and executed by the processor to implement the multi-step prediction method for the zero-crossing point of short-circuit current based on a double-breakpoint circuit breaker as described in any one of claims 1-6.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores at least one instruction, at least one program, code set or instruction set, and the instruction, program, code set or instruction set is loaded and executed by the processor to implement the multi-step prediction method for the zero-crossing point of short-circuit current based on a double-breakpoint circuit breaker as described in any one of claims 1-7.

Images