A power plant relay protection setting optimization method and system based on optimization algorithm

Abstract

The present invention relates to the technical field of relay protection setting optimization, and in particular to a method and system for optimizing relay protection setting values ​​for power plants based on an optimization algorithm. The method comprises acquiring sequential operation data and historical fault data of the power plant; performing data preprocessing on the acquired data; performing feature extraction and fusion on the preprocessed data, and performing multi-scale feature extraction and prediction on the fused data to obtain initial relay protection setting values; optimizing the initial relay protection setting values ​​using a taboo particle swarm algorithm to obtain optimized relay protection setting values. The present invention can more comprehensively search for optimal setting combinations in a complex solution space, so that the optimized setting values ​​can better adapt to the complex and changeable operating conditions of the power system, such as being able to reliably operate under different load levels, fault types, and system topologies.

Description

Technical Field

[0001] The present invention relates to the technical field of relay protection setting optimization, and in particular to a power plant relay protection setting optimization method and system based on an optimization algorithm. Background Art

[0002] Relay protection systems play a vital role in the operation of modern power plants, and the rationality of their settings is directly related to the safe and stable operation of the power system. With the continuous expansion of power system scale, the increasing complexity of operation, and the impact of factors such as the integration of new energy sources, traditional relay protection setting methods are no longer able to meet current needs.

[0003] On the one hand, the time-series operational data and historical fault data generated during power plant operation are massive, high-dimensional, complex, and highly variable. This data contains rich information about system operating status, but is also mixed with a large amount of noise and outliers. Without effective processing and analysis, it is difficult to extract valuable features and patterns for setting relay protection settings.

[0004] On the other hand, the dynamic characteristics of power systems require that relay protection settings be adaptively adjusted according to varying operating conditions and fault conditions. However, traditional setting calculation methods are often based on fixed models and parameters, lacking adaptability and flexibility to real-time system changes. This can lead to false or non-operational protection in certain complex operating conditions, seriously impacting the reliability and stability of power systems.

[0005] Furthermore, existing optimization algorithms for optimizing relay protection settings suffer from slow convergence and a tendency to fall into local optimal solutions, making it difficult to efficiently search for the globally optimal setting combination. Furthermore, there is a lack of comprehensive, accurate, and intelligent methods for evaluating settings, making it difficult to ensure that the optimized settings meet protection performance requirements in various practical operating scenarios.

[0006] Therefore, in order to improve the accuracy, adaptability and reliability of relay protection settings in power plants, there is an urgent need for a relay protection setting optimization method and system that can fully mine and utilize operating data and combine advanced optimization algorithms and intelligent evaluation technologies. Summary of the Invention

[0007] In order to solve the above-mentioned problems, the present invention provides a method and system for optimizing relay protection settings in a power plant based on an optimization algorithm.

[0008] In a first aspect, the present invention provides a method for optimizing relay protection settings in a power plant based on an optimization algorithm, which adopts the following technical solutions:

[0009] A power plant relay protection setting optimization method based on an optimization algorithm includes:

[0010] Obtain power plant timing operation data and historical fault data;

[0011] Perform data preprocessing on the acquired data;

[0012] Perform feature extraction and fusion on the preprocessed data, and perform multi-scale feature extraction and prediction on the fused data to obtain the initial relay protection setting;

[0013] The initial relay protection setting is optimized using the taboo particle swarm optimization algorithm to obtain the optimized relay protection setting.

[0014] The optimized relay protection settings are evaluated using a deep belief network.

[0015] Output the optimization results that meet the evaluation requirements.

[0016] Furthermore, the acquired data is preprocessed, including data cleaning and standardization; wherein data cleaning includes removing noise and outliers, and eliminating abnormal data points with high deviation from the mean, wherein the cleaning conditions are set as follows:

[0017]

[0018] in, and are the mean and standard deviation calculated for the i-th data source in the historical time window:

[0019]

[0020] Where T is the historical time window.

[0021] Furthermore, the pre-processed data is subjected to feature extraction and fusion, including extracting spatial features from the standardized features through CNN, and using GCN to combine the spatiotemporal correlation of multimodal data to generate a fusion feature vector. :

[0022]

[0023] in is the convolution feature of each data source, is the input matrix of the time step t of the multimodal data, consisting of multiple Composition, W is the convolution kernel weight matrix, b is the bias term, A is the adjacency matrix, which represents the spatiotemporal correlation between different features, and the dimension is .

[0024] Furthermore, the fused data is subjected to multi-scale feature extraction and prediction, including the spatiotemporal fusion feature Perform multi-scale convolution operations, using convolution kernels of different sizes Extract features, convolution results Represents the characteristics of the k-th time scale:

[0025]

[0026] in, is the convolution feature matrix of time scale k, is the weight of the kth convolution kernel, which controls the feature extraction of different time ranges. Use a gating mechanism to control the flow of information, expressed as:

[0027]

[0028] in, is the gate value, which represents the importance weight of the feature and is generated by the Sigmoid activation function σ; is the gating unit weight matrix, which represents the learning parameters of the current time scale feature. is the bias term of the gate unit, which is used to adjust the output value. is the activation feature matrix after gating processing.

[0029] Furthermore, the multi-scale feature extraction and prediction of the fused data also includes the activation features at different time scales. Perform weighted fusion to generate the final prediction result :

[0030]

[0031] in, is the final prediction output, which is used to represent the relay protection parameters within a certain period of time in the future. is the weight of the kth time scale, which is determined through model training and is used to balance the contribution of features at different time scales.

[0032] Furthermore, the initial relay protection constant is optimized using the taboo particle swarm algorithm to obtain the optimized relay protection constant, including initializing the particle swarm and the taboo table based on the initial relay protection constant, updating the individual optimum and the global optimum of the particles based on the fitness function and updating the particle speed and position, performing taboo processing on the new position of the particle based on the taboo table, and obtaining the optimized relay protection constant through updating and iteration.

[0033] Furthermore, the use of a deep belief network to evaluate the optimized relay protection settings includes optimizing the deep belief network structure using random search, and training each layer of the deep belief network structure using a restricted Boltzmann machine, wherein the training step size and learning rate are dynamically adjusted through an improved contrast divergence algorithm to accelerate the convergence speed, and a dynamic adjustment mechanism for the evaluation threshold is established based on statistical analysis of power plant data, and qualified relay protection settings are output through relay protection setting evaluation.

[0034] In the second aspect, a power plant relay protection setting optimization system based on an optimization algorithm includes:

[0035] The data acquisition module is configured to acquire the power plant's sequential operation data and historical fault data;

[0036] The preprocessing module is configured to perform data preprocessing on the acquired data;

[0037] The feature module is configured to extract and fuse features of the preprocessed data, and perform multi-scale feature extraction and prediction on the fused data to obtain an initial relay protection setting;

[0038] The optimization module is configured to optimize the initial relay protection setting value by using a taboo particle swarm optimization algorithm to obtain an optimized relay protection setting value;

[0039] The evaluation module is configured to evaluate the optimized relay protection settings using a deep belief network;

[0040] The output module is configured to output an optimized result that is evaluated as qualified.

[0041] In a third aspect, the present invention provides a computer-readable storage medium storing a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor of a terminal device, for a power plant relay protection setting optimization method based on an optimization algorithm.

[0042] In a fourth aspect, the present invention provides a terminal device comprising a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, and the instructions are suitable for being loaded and executed by the processor to implement the method for optimizing the constant value of power plant relay protection based on an optimization algorithm.

[0043] In summary, the present invention has the following beneficial technical effects:

[0044] 1. This invention removes noise and outliers by comprehensively collecting and meticulously preprocessing power plant sequential operation data and historical fault data, more accurately reflecting the system's true operating status and providing a reliable foundation for subsequent constant value calculations. It also optimizes the initial relay protection constants based on a taboo particle swarm optimization algorithm. This algorithm, based on the traditional particle swarm optimization algorithm, introduces a taboo search mechanism, effectively avoiding the problem of being trapped in a local optimal solution. It can more comprehensively search for the optimal constant value combination in a complex solution space, enabling the optimized constant values ​​to better adapt to the complex and changing operating conditions of the power system, such as reliable operation under different load levels, fault types, and system topologies.

[0045] 2. Utilizing CNN to extract spatial features and GCN to combine spatiotemporal correlation to generate fused feature vectors, as well as multi-scale feature extraction and prediction technology, we fully exploit the multi-dimensional information in the data. Compared with traditional single-scale or simple feature extraction methods, we can more accurately capture subtle changes in the power system under different operating conditions, thereby obtaining more accurate initial relay protection settings and reducing setting errors caused by insufficient feature extraction.

[0046] 3. During the multi-scale feature extraction and prediction process, by weighted fusion of features at different time scales and determining the weight of each scale based on model training, the generated relay protection constants can comprehensively consider short-term and long-term system operation trends, further enhancing the constants' adaptability to system dynamic changes and ensuring timely and accurate response to faults under various operating conditions.

[0047] 4. The optimized relay protection settings are evaluated using a deep belief network, and the network structure is optimized using random search. This approach can automatically learn the complex mapping relationship between the settings and the system operating status. Compared with traditional evaluation methods based on fixed rules or simple experience, it has stronger generalization ability and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a schematic diagram of a power plant relay protection setting optimization method based on an optimization algorithm according to Example 1 of the present invention. DETAILED DESCRIPTION

[0049] The present invention will be further described in detail below with reference to the accompanying drawings.

[0050] Example 1

[0051] Reference Figure 1 , a power plant relay protection setting optimization method based on an optimization algorithm in this embodiment includes:

[0052] Obtain power plant timing operation data and historical fault data;

[0053] Perform data preprocessing on the acquired data;

[0054] Perform feature extraction and fusion on the preprocessed data, and perform multi-scale feature extraction and prediction on the fused data to obtain the initial relay protection setting;

[0055] The initial relay protection setting is optimized using the taboo particle swarm optimization algorithm to obtain the optimized relay protection setting.

[0056] The optimized relay protection settings are evaluated using a deep belief network.

[0057] Output the optimization results that meet the evaluation requirements.

[0058] Specifically:

[0059] S1. Collect various data related to relay protection from the power plant's monitoring system, historical operating data, and power equipment parameter database, including line current, voltage, power, equipment rated parameters, fault records, etc.

[0060] Multimodal data collected from different sources can be defined as ,in: Represents the original time series data collected by the i-th data source, each data source The time series data is represented as: ,in, is the jth eigenvalue collected at time t. All collected data D are stored in the time series database for further processing.

[0061] S2. Data preprocessing and fusion

[0062] Before data fusion, the raw data is first cleaned, standardized and feature extracted to ensure data quality and consistency.

[0063] Data cleaning: remove noise and outliers, and remove abnormal data points that deviate greatly from the mean. Set the cleaning conditions as follows:

[0064]

[0065] in, and are the mean and standard deviation calculated for the i-th data source in the historical time window:

[0066]

[0067] Where T is the historical time window.

[0068] Standardization: Mapping the values ​​of each data source to a unified dimension, that is, the cleaned data , normalized to the data range [0,1] :

[0069]

[0070] in, and are the minimum and maximum values ​​of the time window.

[0071] Feature extraction and fusion: The standardized features are used to extract spatial features through CNN, and GCN is used to combine the spatiotemporal correlation of multimodal data to generate a fused feature vector. :

[0072]

[0073] in is the convolution feature of each data source, is the input matrix of the time step t of the multimodal data, consisting of multiple W is the convolution kernel weight matrix, b is the bias term. A is the adjacency matrix, which represents the spatiotemporal correlation between different features. Its dimension is .

[0074] S3. Multi-scale feature extraction and intelligent prediction

[0075] This module is responsible for fusing features from time and space The multi-scale dynamic features are extracted from the model and the relay protection settings are predicted with high accuracy. By introducing the multi-scale gated Tanh unit (M-GTU), the model's ability to capture features at different time scales is enhanced.

[0076] Multi-scale feature extraction: spatiotemporal fusion features Perform multi-scale convolution operations, using convolution kernels of different sizes Extract features. Convolution results Represents the characteristics of the k-th time scale:

[0077]

[0078] in, is the convolution feature matrix of time scale k, is the weight of the kth convolution kernel, which controls the feature extraction of different time ranges.

[0079] Gating processing mechanism: features at each time scale Use a gating mechanism to control the flow of information. The calculation formula is:

[0080]

[0081] in, is the gate value, which represents the importance weight of the feature and is generated by the Sigmoid activation function σ. is the gating unit weight matrix, which represents the learning parameters of the current time scale feature. is the bias term of the gate unit, which is used to adjust the output value. is the activation feature matrix after gating processing.

[0082] Multi-scale feature fusion and prediction: Activation features at different time scales Perform weighted fusion to generate the final prediction result :

[0083]

[0084] in, It is the final prediction output, used to represent the preliminary relay protection setting. is the weight of the kth time scale, which is determined through model training and is used to balance the contribution of features at different time scales.

[0085] S4. Determine the size of the particle swarm: Determine the number of particles in the particle swarm based on the complexity of the problem and computing resources. n , usually between 20 and 100.

[0086] Set the position and speed range: According to the actual value range of the relay protection setting, set the upper and lower limits of the particle position xi and speed vi to ensure that the particles search within a reasonable range.

[0087] Initialize the tabu table: Create an empty tabu table TabuList to record the solutions that have been visited, and set the tabu length TL, which is the maximum number of solutions recorded in the tabu table.

[0088] Determine the termination condition of the algorithm: You can set the maximum number of iterations MaxIter or reach a certain accuracy requirement as the termination condition of the algorithm.

[0089] Generate initial particle swarm

[0090] The initial position Xi of particles is randomly generated within the set position range, and the position of each particle represents a set of relay protection settings.

[0091] Randomly initialize the particle's velocity Vi, which determines the particle's moving direction and step size in the search space.

[0092] Calculating the fitness function

[0093] Define the fitness function f(x) to evaluate the quality of each set of relay protection settings. Usually, the fitness function can be constructed based on the performance indicators of relay protection, such as the accuracy, speed, and reliability of protection action.

[0094] For each particle i, calculate the fitness value f(xi) corresponding to its current position xi.

[0095] Update the individual optimum and global optimum of particles

[0096] Individual optimal update: For each particle i, compare the fitness value f(xi) of its current position with the fitness value f(pi) of the individual's historical optimal position pi. If f(xi) < f(pi), update the individual optimal position pi = xi.

[0097] Global optimal update: Among the individual optimal positions of all particles, find the position with the minimum (or maximum, depending on the optimization goal) fitness value as the global optimal position g.

[0098] Iterative process of taboo particle swarm optimization algorithm

[0099] Update speed and position: Update the particle's speed vi and position xi according to the basic formula of the particle swarm algorithm:

[0100]

[0101] Among them, w is the inertia weight, C1 and C2 are learning factors, r1 and r2 are random numbers between 0 and 1, and t is the number of iterations.

[0102] After updating the particle position, the newly generated position is tabooed. Check whether the new position xi(t+1) is in the taboo list (TabuList). If it is, adjust it according to a specific strategy, such as randomly generating a new position or fine-tuning the current position within a certain range, until a feasible position is generated that is not in the taboo list.

[0103] Add the newly generated position xi(t+1) to the taboo list TabuList. If the length of the taboo list exceeds the taboo length TL, delete the earliest added record to keep the length of the taboo list unchanged.

[0104] Recalculate the fitness function: For the updated particle position xi(t+1), recalculate its fitness value f(xi(t+1)).

[0105] Update individual optimal and global optimal: According to the above method, update the individual optimal position pi and global optimal position g of the particle.

[0106] Check whether the termination conditions of the algorithm are met, such as reaching the maximum number of iterations MaxIter or the fitness value of the global optimal position does not change significantly in multiple consecutive iterations.

[0107] If the termination condition is met, the relay protection setting corresponding to the global optimal position g is output as the optimized result; otherwise, it returns to continue the iteration.

[0108] (1) Constructing a deep belief network (DBN) by building a deep belief network structure with multiple hidden layers to fully learn the complex intrinsic characteristics and potential patterns of the data. The number of hidden layers and nodes is dynamically adjusted according to the complexity of the data and the scale of the problem. In the initial stage, a smaller network structure is determined through experiments, such as 3 hidden layers with 100, 50, and 20 nodes per layer. Then, an automatic hyperparameter random search optimization algorithm is used to further optimize the network structure to find the optimal hidden layer configuration to balance the model's fitting ability and computational complexity.

[0109] (2) Based on improved RBM training: In the training of each layer of the restricted Boltzmann machine (RBM), an improved contrastive divergence algorithm, namely the fast continuous contrastive divergence algorithm (FPC), is used. This algorithm accelerates the convergence of the RBM by dynamically adjusting the training step size and learning rate, and improves its ability to learn complex data distributions. During the training process, the momentum term is introduced to enable the model to better utilize historical gradient information when updating parameters, avoid falling into local optimal solutions, and enhance the stability and robustness of the model.

[0110] (3) When training RBMs layer by layer, an adaptive layer-by-layer training strategy is adopted. First, principal component analysis (PCA) is performed on the input data. The initial learning rate and number of training steps for each layer of RBM are determined based on the principal component contribution rate of the data. During the training process, the reconstruction error and information entropy of the output features of each layer of RBM are monitored, and the training parameters of subsequent layers are dynamically adjusted to ensure that each layer can effectively learn the important features of the data and avoid excessive compression or loss of feature information.

[0111] (4) Add an output layer on top of the DBN and use a combination of the Connectionist Temporal Classification (CTC) loss function and the traditional Mean Squared Error (MSE) loss function as the objective function for supervised fine-tuning. The CTC loss function is often used to handle the classification of sequence data and can effectively deal with the time series correlation and label inconsistency problems that may exist in relay protection setting value evaluation. The MSE loss function is used to directly optimize the error between the predicted setting value and the true setting value. By adjusting the weights of these two loss functions, the performance of the model in terms of classification accuracy and setting value prediction accuracy can be balanced.

[0112] (5) During the fine-tuning process, an attention mechanism is introduced to enable the model to automatically focus on the key features and information in the input data. By calculating the attention weights between the input features and the hidden layer states, the model can more effectively focus on the parts closely related to the relay protection setting evaluation and ignore irrelevant noise information, thereby improving the accuracy and pertinence of the evaluation. At the same time, an adaptive learning rate adjustment strategy (such as AdaGrad, RMSProp, or Adam optimizer) is used to automatically adjust the learning rate of each parameter based on the parameter update history, accelerating the convergence speed of the model and, to a certain extent, alleviating the gradient vanishing or gradient exploding problems.

[0113] (6) Use k-fold cross-validation (e.g., k = 5 or 10) to evaluate and select multiple trained DBN models. Divide the dataset into k mutually exclusive subsets, select one of the subsets as the validation set each time, and the remaining subsets as the training set. Train and validate the model k times. By calculating the average evaluation metrics (such as accuracy, recall, F1 value, root mean square error, etc.) of each model on the validation set, select the model with the best performance as the final evaluation model.

[0114] (7) Based on the actual operating conditions of the power plant and statistical analysis of historical data, a dynamic adjustment mechanism for the evaluation threshold is established. For example, under different operating conditions (such as peak load period, valley load period, equipment maintenance period, etc.), the reasonable range of relay protection setting values ​​may change. Therefore, it is necessary to dynamically adjust the evaluation threshold based on real-time operating condition information to make the evaluation results more consistent with the actual situation and operating requirements. At the same time, using online learning and incremental learning methods, the parameters and thresholds of the model are continuously updated so that it can adapt to changes in the operating status of the power system and emerging fault modes, thereby realizing adaptive relay protection setting evaluation.

[0115] By using the deep belief network to evaluate the relay protection settings optimized by the taboo particle swarm algorithm, various factors and complex situations can be considered more accurately and comprehensively, the reliability and practicality of the evaluation results can be improved, and strong support can be provided for the optimization and safe and stable operation of the power plant's relay protection system.

[0116] Example 2

[0117] This embodiment provides a power plant relay protection setting optimization system based on an optimization algorithm, including:

[0118] The data acquisition module is configured to acquire the power plant's sequential operation data and historical fault data;

[0119] The preprocessing module is configured to perform data preprocessing on the acquired data;

[0120] The feature module is configured to extract and fuse features of the preprocessed data, and perform multi-scale feature extraction and prediction on the fused data to obtain an initial relay protection setting;

[0121] The optimization module is configured to optimize the initial relay protection setting value by using a taboo particle swarm optimization algorithm to obtain an optimized relay protection setting value;

[0122] The evaluation module is configured to evaluate the optimized relay protection settings using a deep belief network;

[0123] The output module is configured to output an optimized result that is evaluated as qualified.

[0124] A computer-readable storage medium stores a plurality of instructions, wherein the instructions are suitable for being loaded by a processor of a terminal device and executing the method described.

[0125] A terminal device includes a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, and the instructions are suitable for being loaded by the processor and executing the method described.

[0126] The above are all preferred embodiments of the present invention, and are not intended to limit the scope of protection of the present invention. Therefore, any equivalent changes made based on the structure, shape, and principle of the present invention should be included in the scope of protection of the present invention.

Claims

1. A power plant relay protection setting optimization method based on an optimization algorithm, characterized in that: include: Obtain power plant timing operation data and historical fault data; Perform data preprocessing on the acquired data; Perform feature extraction and fusion on the preprocessed data, and perform multi-scale feature extraction and prediction on the fused data to obtain the initial relay protection setting; The initial relay protection setting is optimized using the taboo particle swarm optimization algorithm to obtain the optimized relay protection setting. The optimized relay protection settings are evaluated using a deep belief network. Output the optimization results that meet the evaluation requirements; The data preprocessing includes data cleaning and standardization. The data cleaning includes removing noise and outliers, and eliminating abnormal data points with high deviation from the mean. The cleaning conditions are set as follows: in, is the jth eigenvalue collected at time t, and are the mean and standard deviation calculated for the i-th data source in the historical time window: Where T is the historical time window; The pre-processed data is subjected to feature extraction and fusion, including extracting spatial features from the standardized features through CNN, and using GCN to combine the spatiotemporal correlation of multimodal data to generate a fusion feature vector. : in is the convolution feature of each data source, is the input matrix of the time step t of the multimodal data, consisting of multiple Composition, W is the convolution kernel weight matrix, b is the bias term, A is the adjacency matrix, which represents the spatiotemporal correlation between different features, and the dimension is ; The multi-scale feature extraction and prediction of the fused data, including the spatiotemporal fusion feature Perform multi-scale convolution operations, using convolution kernels of different sizes Extract features, convolution results Represents the characteristics of the k-th time scale: in, is the convolution feature matrix of time scale k, is the weight of the kth convolution kernel, which controls the feature extraction of different time ranges. Use a gating mechanism to control the flow of information, expressed as: in, is the gate value, which represents the importance weight of the feature and is generated by the Sigmoid activation function σ; is the gate unit weight matrix, which represents the learning parameters of the current time scale feature; is the bias term of the gate control unit, which is used to adjust the output value; is the activation feature matrix after gating processing; The multi-scale feature extraction and prediction of the fused data also includes the activation features at different time scales. Perform weighted fusion to generate the final prediction result : in, is the final prediction output, which is used to represent the relay protection parameters within a certain period of time in the future. is the weight of the kth time scale, determined through model training, and used to balance the contributions of features at different time scales; The method uses a taboo particle swarm algorithm to optimize the initial relay protection constant to obtain the optimized relay protection constant, including initializing a particle swarm and a taboo table based on the initial relay protection constant, updating the individual optimum and the global optimum of the particles based on the fitness function and updating the particle speed and position, performing taboo processing on the new position of the particle based on the taboo table, and obtaining the optimized relay protection constant through updating and iteration.

2. The power plant relay protection setting optimization method based on the optimization algorithm according to claim 1 is characterized in that: The method uses a deep belief network to evaluate the optimized relay protection settings, including optimizing the deep belief network structure using random search, and training each layer of the deep belief network structure using a restricted Boltzmann machine. The method uses an improved contrast divergence algorithm to dynamically adjust the training step size and learning rate to accelerate the convergence speed, establishes a dynamic adjustment mechanism for the evaluation threshold based on statistical analysis of power plant data, and outputs qualified relay protection settings through relay protection setting evaluation.

3. A power plant relay protection setting optimization system based on an optimization algorithm, executing a power plant relay protection setting optimization method based on an optimization algorithm as claimed in claim 1, characterized in that: include: The data acquisition module is configured to acquire the power plant's sequential operation data and historical fault data; The preprocessing module is configured to perform data preprocessing on the acquired data; The feature module is configured to extract and fuse features of the preprocessed data, and perform multi-scale feature extraction and prediction on the fused data to obtain an initial relay protection setting; The optimization module is configured to optimize the initial relay protection setting value by using a taboo particle swarm optimization algorithm to obtain an optimized relay protection setting value; The evaluation module is configured to evaluate the optimized relay protection settings using a deep belief network; The output module is configured to output an optimized result that is evaluated as qualified.

4. A computer-readable storage medium storing a plurality of instructions, characterized in that: The instructions are suitable for being loaded by a processor of a terminal device and executing the method according to claim 1 .

5. A terminal device comprising a processor and a computer-readable storage medium, wherein the processor is configured to implement various instructions; and the computer-readable storage medium is configured to store a plurality of instructions, wherein: The instructions are suitable for being loaded by a processor and for executing the method according to claim 1 .

Images