Short-term wind power prediction method

Abstract

The invention discloses a short-term wind power prediction method, and the method comprises the steps: S1, collecting historical wind power data, and carrying out the preprocessing of the data, S2, improving a frost ice optimization algorithm, and extracting the multi-frequency-domain features of the wind power, S3, constructing an LSTM model, and improving a fishing algorithm, and S4, determining the optimal learning rate and penalty factor of the LSTM model based on the S3, and obtaining an ECFOA-LSTM network model, and S5, based on the IMF component set obtained in the S2, obtaining a wind power prediction value. According to the invention, the frost ice optimization algorithm and the fishing algorithm are improved, the optimal learning rate and penalty factor of the model are determined through the improved fishing algorithm, and finally the ECFOA-LSTM network model is obtained.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power prediction, and particularly to a short-term wind power prediction method. Background Art

[0002] Wind power plays an irreplaceable role in addressing environmental protection issues and energy crises. However, its strong non-stationarity and intermittency affect the stable operation of the power grid, power generation efficiency, and power quality. Accurate wind power prediction technology is crucial for the stable operation and economic dispatching of the power system, and helps wind farms formulate scientific and reasonable control strategies, improve the utilization rate of wind energy. From the perspective of the usage model, prediction methods can be mainly divided into physical models, statistical models, and combined models. In recent years, due to its high prediction accuracy, the combined prediction model has gradually replaced the single prediction model and become the mainstream research direction. The combined prediction model applies intelligent optimization algorithms to the field of wind power prediction, and further improves the prediction accuracy by optimizing the parameters or structure of the wind speed prediction model.

[0003] There are bottlenecks in existing short-term wind power prediction technologies: Variational Mode Decomposition (VMD) lacks adaptability, the signal-to-noise ratio of the signal source is low, frequency domain aliasing occurs, and it is easy to over-filter or under-filter, resulting in inaccurate estimation of the number of signal sources; when the separation algorithm decomposes the signal source, there are problems such as incorrect filtering of the source signal, inaccurate estimation of the number of signal sources, and large time-frequency domain distribution; the fitting ability of the LSTM model depends on hyperparameters, and traditional parameter tuning has low efficiency and poor accuracy; the real-time monitoring and prediction method based on sensor data has high costs, is limited by the accuracy and reliability of the equipment, has low economy, and limited promotion prospects.

[0004] In summary, the existing technologies mainly show problems such as low prediction accuracy and poor economy in short-term wind power prediction, and it is difficult to effectively support the further development of wind power generation. Summary of the Invention

[0005] The purpose of the present invention is to provide a short-term wind power prediction method to solve the problem of excessive error in the LSTM prediction model in predicting wind power generation in the prior art.

[0006] The technical solution adopted by the present invention is a short-term wind power prediction method, and the steps include: Step S1, collect historical wind power data and preprocess the data; Step S2, use an improved frost ice optimization algorithm to extract multi-frequency domain features of wind power; Step S3, construct an LSTM model and an improved fishing algorithm; Step S4, based on S3, determine the optimal learning rate and penalty factor of the LSTM model to obtain an ECFOA-LSTM network model; Step S5: Based on the set of IMF components obtained in S2, obtain the wind power prediction value.

[0007] The specific steps of S2 are as follows: S21: Improve the frost ice optimization algorithm by changing the cosine transform of the frost ice factor to a double transform of sine and cosine as follows: ; where rand represents a random number obeying a normal distribution within (0, 1), Riter represents the current iteration number of the frost ice optimization algorithm, Max_Riter represents the maximum iteration number of the frost ice optimization algorithm, W represents the parameter of the algorithm, which is taken as 5 in the present invention, and RimeFactor represents the frost ice factor, which determines the performance of the frost ice optimization algorithm; S22: Optimize the penalty coefficient and the number of modes in the variational mode decomposition through ERIME, and the formula is as follows: ; where E represents the coefficient of determination using the update method, Rimepop represents the updated penalty coefficient and mode combination, best_rime represents the current best penalty coefficient and mode combination with the minimum sample entropy, represents the upper limit of the penalty coefficient and the number of modes, represents the lower limit of the penalty coefficient and the number of modes, rand and represent random numbers obeying a normal distribution within (0, 1); S23: Take the minimum value of the sample entropy as the objective function, perform variational mode decomposition on the wind power data in the training set and the test set, and obtain the set of intrinsic mode function components of the wind power. Each IMF component represents the local characteristics of different frequency domains, and continuously optimize the number of modes K and the penalty coefficient through the minimum sample entropy to find the best IMF component. The smaller the sample entropy, the better the IMF component. The specific formula is as follows: ; where SE represents the sample entropy, K represents the number of modes, represents the penalty coefficient, arg min represents the minimization operation, represents the index of the sample entropy.

[0008] The specific steps to improve the fishing algorithm in S3 are as follows: S31: Initialize the population size, and the formula is as follows: ; where represents the group of learning rates and penalty factors at the The position in the d-dimensional search space, represents the upper bounds of the learning rate and the penalty factor in the d-dimensional search space, represents the lower bounds of the learning rate and the penalty factor in the d-dimensional search space, and rand represents a random number obeying the normal distribution within (0~1); S32: Divide the exploration stage into an independent search mode and a group fishing mode. The calculation formula for the capture rate is as follows: ; where, represents the capture rate, EFs represents the current evaluation quantity, and MaxEFs represents the maximum evaluation quantity; S33: Improve the independent search mode of the fishing algorithm using the difference idea, and update the learning rate and the penalty factor at the current position. The specific formulas are as follows: ; ; In the formula, Exp is the empirical value parameter, and its value range is (-1~1). represents the worst fitness value (i.e., the maximum error, the larger the worse) after the th position update, represents the best fitness value (i.e., the minimum root mean square error RMSE) after the th position update, represents the fitness value of the learning rate and the penalty factor, represents the th iteration position corresponding to the fitness value of the learning rate and the penalty factor, represents the exploration range, represents the Euclidean distance between the learning rate and the penalty factor, EFs represents the current evaluation quantity, MaxEFs represents the maximum evaluation quantity, represents the iteration number of the learning rate and the penalty factor, represents the th iteration, the position of the th group of learning rate and penalty factor in the d-dimensional search space, represents the th iteration, the position of the th group of learning rate and penalty factor in the d-dimensional search space, represents the th iteration position corresponding to the learning rate and the penalty factor at in the d-dimensional search space, is a random number following a normal distribution between (0~1); s represents a d-dimensional random unit vector used to determine the main travel direction and distance; represents a random number following a normal distribution within (0~2), represents a random number following a normal distribution within (1~3), and rand represents a random number following a normal distribution within (0~1); S34. In the simulation of the group fishing mode, 3 to 4 groups of learning rates and penalty factors are sequentially selected to form a subgroup , the subgroup obtains the discrete Laplace operator through discrete Laplace transform, and the specific formula is as follows: ; ; In the formula, represents the target point enclosed by the subgroup , represents the th iteration, and the position of the th group of learning rate and penalty factor in the subgroup in the -dimensional search space, represents the th iteration, and the position of the th group of learning rate and penalty factor in the subgroup in the -dimensional search space; represents the speed of the fisherman approaching the center, with a value range of (0~1), is the offset of movement, with a value range of (-1~1), gradually decreases as EFs increases; EFs represents the current evaluation quantity, MaxEFs represents the maximum evaluation quantity, laplace represents the discrete Laplace operator, and A represents the matrix that randomly arranges the population, represents the th iteration, and the position of the learning rate and penalty factor in the subgroup , and mean represents the mean of the parameters, that is, the center position of the group; S35. In the simulation of the large exploitation stage of the fishing algorithm, an improved diagonal form is used for update, and the formula is as follows: ; ; In the formula, represents the variance, EFs represents the current evaluation quantity, and MaxEFs represents the maximum evaluation quantity, represents the The position of the learning rate and penalty factor of the group after iteration, is the global optimal position, GD is the Gaussian distribution function, represents a random number subject to a normal distribution within (1-3), which is used to distribute the learning rate and penalty factor within three ranges; represents the average value matrix of each dimension at the center of the learning rate and penalty factor. diag(A) means taking the values on the diagonal after randomly arranging the population as the update condition for this round of iteration. mean represents the mean value of the parameter, and rand represents a random number subject to a normal distribution within (0-1).

[0009] The specific steps of the said S5 are as follows: S51, Take the first 70% of the IMF component set as the training set and input it into the ECFOA-LSTM network model to train the model; S52, Take the remaining 30% of the IMF components in the IMF component set as the prediction set and input it into the trained ECFOA-LSTM network model to obtain the predicted values corresponding to the IMF components; S53, Add up the predicted values corresponding to all IMF components in the test set, and the sum is the predicted value of wind power.

[0010] The beneficial effects of the present invention are: The present invention breaks through the bottleneck of wind power prediction of a single LSTM prediction model. Compared with the existing single LSTM prediction model, it has high prediction accuracy and good economy, and effectively supports the reliable prediction of wind power generation.

[0011] Based on the VMD optimization of ERIME, the present invention significantly improves the signal decomposition effect.

[0012] The present invention uses ECFOA to optimize the LSTM model, which has better model performance.

[0013] The present invention adopts a stochastic improved ice cream optimization algorithm, which enhances the randomness of the algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0015] Figure 1 is the step flow chart of the present invention.

[0016] Figure 2It is a comparison chart of the prediction results of each model.

[0017] Figure 3(a) is a comparison chart of the effects of the frost ice optimization algorithm before and after improvement on the test function F1.

[0018] Figure 3(b) is a comparison chart of the effects of the frost ice optimization algorithm before and after improvement on the test function F2.

[0019] Figure 3(c) is a comparison chart of the effects of the frost ice optimization algorithm before and after improvement on the test function F3.

[0020] Figure 3(d) is a comparison chart of the effects of the frost ice optimization algorithm before and after improvement on the test function F4.

[0021] Figure 4 It is a comparison chart of the effects of the frost ice optimization algorithm before and after improvement on optimizing the VMD sample entropy value.

[0022] Figure 5(a) is a comparison chart of the effects of the fishing algorithm before and after improvement on the test function F1.

[0023] Figure 5(b) is a comparison chart of the effects of the fishing algorithm before and after improvement on the test function F2.

[0024] Figure 5(c) is a comparison chart of the effects of the fishing algorithm before and after improvement on the test function F3.

[0025] Figure 5(d) is a comparison chart of the effects of the fishing algorithm before and after improvement on the test function F4. Specific implementation manners

[0026] 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 the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention. Embodiment

[0027] An embodiment of the present invention provides a short-term wind power prediction method, and the flow chart is as Figure 1 shown, and the steps include: Step S1, collect historical wind power data, preprocess the data, construct a short-term wind power prediction data set, and then divide the short-term wind power prediction data set into a training set and a test set according to a quantity ratio of 7:3. The short-term wind power prediction data set includes power generation, wind speed, wind direction, and temperature information, and the sampling frequency is 1 hour (a total of 24 data points throughout the day). Based on the actual minimum and maximum power values of the historical data, the original power data is mapped to the [0,1] interval using the min-max normalization method.

[0028] Step S2, extract the multi-frequency domain features of wind power by the improved frost ice optimization algorithm (ERIME), and the specific steps are as follows: S21, improve the frost ice optimization algorithm, and improve the cosine transform of the frost ice factor to a positive and cosine double transform to enhance the algorithm performance to escape the local optimal solution and expand the search area of the algorithm, specifically as follows: ; where rand represents a random number subject to a normal distribution within (0~1), Riter represents the current iteration number of the frost ice optimization algorithm, Max_Riter represents the maximum iteration number of the frost ice optimization algorithm, W represents the parameter of the algorithm, which is taken as 5 in the present invention, and RimeFactor represents the frost ice factor, which determines the performance of the frost ice optimization algorithm.

[0029] S22, optimize the penalty coefficient and the number of modes in the variational mode decomposition (VMD) by ERIME to improve the accuracy of signal decomposition. In the prior art, the determination coefficient of the update method of the frost ice optimization algorithm is usually determined according to the current iteration number and the maximum iteration number. In the present invention, a random number within 0~1 is selected as the determination coefficient of the update method, and two independent random numbers rand and are generated in each round of iteration. The formula is as follows: ; where E represents the determination coefficient of the update method, Rimepop represents the updated penalty coefficient and mode combination, best_rime represents the current best penalty coefficient and mode combination with the minimum sample entropy, represents the upper limit of the penalty coefficient and the number of modes, represents the lower limit of the penalty coefficient and the number of modes, and rand and represent random numbers subject to a normal distribution within (0~1); The comparison between ERIME and the standard RIME retesting function is shown in Figure 3 (the maximum number of iterations is 1000, and the population size is 30). Among them, a~d are the two algorithms for solving the functions F1 (Shifted Sphere Function), F2 (Shifted Schwefel's Problem 1.2), F3 (Shifted Rotated High Conditioned Elliptic Function), and F4 (Shifted Schwefel's Problem 1.2 with Noise in Fitness Function). Among them, the left figure is the spatial surface of the corresponding function, and the right figure is the iterative convergence accuracy of ERIME and the standard RIME. The four functions have unimodal convergence, complex variable dependence, rotational inseparability, and noise robustness. Through these four functions, the comprehensive performance of the algorithm under ideal conditions, complex coupling, and real noise scenarios can be comprehensively evaluated, providing a key benchmark for improving algorithm design. It can be seen from Figure 3 that as the number of iterations increases, ERIME has a faster convergence speed and higher convergence accuracy.

[0030] S23. Taking the minimum sample entropy as the objective function, variational mode decomposition is performed on the wind power data on the training set and the test set to obtain the set of intrinsic mode function (IMF) components of the wind power. Each IMF component characterizes the local features of different frequency domains, and the mode number K and the penalty coefficient are continuously optimized through the minimum sample entropy , to find the best IMF component. The smaller the sample entropy, the better the IMF component. The specific formula is as follows: ; Among them, SE represents the sample entropy, K represents the mode number, represents the penalty coefficient, arg min represents the minimization operation, represents the index of the sample entropy.

[0031] The results of ERIME and RIME optimizing VMD are as Figure 2 shown. It can be seen through Figure 2 that compared with RIME, ERIME of the present invention has better optimization ability and maintains better ability in 20 iterations, and the sample entropy is less than the solution result of RIME.

[0032] Step S3, construct an LSTM model, and adopt an improved fish swarm optimization algorithm (ECFOA). By using the improved fish swarm optimization algorithm, find the optimal learning rate and penalty factor to improve the convergence speed and generalization performance of the model, and avoid gradient explosion or gradient disappearance caused by too large or too small learning rate and penalty factor, which may lead to distortion of prediction data and thus render the prediction results unusable. The LSTM model is a known model in the prior art, including an input layer, an LSTM layer, and an output layer. The input layer is used to receive the first data set, the LSTM layer is used to solve the long-term dependence relationship in the data set, and the output layer is used to predict the short-term wind power. The specific steps of the improved fish swarm optimization algorithm are as follows: S31, initialize the population size, and the formula is as follows: ; where, represents the position of the th group of learning rate and penalty factor in the -dimensional search space, represents the upper bound of the learning rate and penalty factor in the -dimensional search space, represents the lower bound of the learning rate and penalty factor in the th-dimensional search space, and rand represents a random number that follows a normal distribution within (0~1).

[0033] The fish swarm optimization algorithm is divided into an exploration stage and a development stage. In the exploration stage, the fish swarm optimization algorithm conducts a global search. When entering the development stage, the fish swarm optimization algorithm conducts a local search.

[0034] S32, divide the exploration stage into an independent search mode and a group fishing mode. The conversion between the independent search mode and the group fishing mode is determined according to the capture rate , and the calculation formula of the capture rate is as follows: ; where, represents the capture rate, EFs represents the current evaluation quantity, and MaxEFs represents the maximum evaluation quantity.

[0035] When the capture rate tends to 1, the fishermen adopt the independent search mode. When the capture rate tends to 0, the fishermen adopt the group fishing mode. Use the random number for simulation. When , adopt the independent search mode, , adopt the group fishing mode.

[0036] S33, use the difference idea to improve the independent search mode of the fish swarm optimization algorithm, and update the learning rate and penalty factor of the current position. The specific formula is as follows: ; ; In the formula, Exp is an empirical value parameter, and its value range is (-1, 1). represents the worst fitness value (i.e., the largest error, the larger the worse) after the th position update, represents the best fitness value (i.e., the smallest root mean square error RMSE) after the th position update, represents the fitness value of the learning rate and the penalty factor, represents the position after the th iteration corresponding to the fitness value of the learning rate and the penalty factor at represents the exploration range, represents the Euclidean distance between the learning rate and the penalty factor, EFs represents the current number of evaluations, and MaxEFs represents the maximum number of evaluations. represents the number of iterations of the learning rate and the penalty factor, represents the th iteration, the position of the th group of learning rate and penalty factor in the dimensional search space, represents the th iteration, the position of the th group of learning rate and penalty factor in the dimensional search space, represents the position after the th iteration corresponding to the learning rate and the penalty factor at in the dimensional search space, is a random number obeying the normal distribution between (0~1); s represents a d-dimensional random unit vector, which is used to determine the main traveling direction and distance; represents a random number obeying the normal distribution within (0~2),

[0037] S34. In the simulated population fishing mode, 3 to 4 groups of learning rates and penalty factors are sequentially selected to form a subgroup , and the subgroup obtains the discrete Laplace operator through the discrete Laplace transform, and then obtains a new position, enabling the algorithm to escape from the local optimum and expand the search range. The specific formula is as follows: ; In the formula, Indicates a subgroup The surrounded target point Indicates the subgroup at the th iteration The position of the th group of learning rate and penalty factor in the th dimensional search space Indicates the subgroup at the th iteration The position of the th group of learning rate and penalty factor in the th dimensional search space; Indicates the speed at which the fisherman approaches the center, with a value range of (0~1), is the offset for movement, with a value range of (-1~1), and gradually decreases as the number of EFs increases; EFs represents the current number of evaluations, MaxEFs represents the maximum number of evaluations, laplace represents the discrete Laplace operator, A represents the matrix for randomly arranging the population, Indicates the position of the learning rate and penalty factor in the subgroup at the th iteration, mean represents the mean of the parameters, that is,

[0038] S35, in the simulated large exploitation stage of the fishing algorithm, an improved diagonal form is used for updating to enhance the memory ability of the algorithm and accelerate the algorithm convergence. The formula is as follows: ; In the formula, represents the variance, EFs represents the current number of evaluations, MaxEFs represents the maximum number of evaluations, Indicates the position of the th group of learning rate and penalty factor after the th iteration, is the global optimal position, GD is the Gaussian distribution function, represents a random number obeying the normal distribution within (1~3), which is used to distribute the learning rate and penalty factor in three ranges;

[0039] Under the experimental settings of a maximum number of iterations of 1000 and a population size of 30, the performance comparison between the ECFOA of the present invention and the standard CFOA on test functions is shown in Figure 5. Four typical functions are compared in the figure. Among them, the left figures in a~d show the spatial surface forms of each function, and the right figures present the iterative convergence curves of the two algorithms. The experimental results show that as the number of iterations increases, the ECFOA exhibits a faster convergence speed and higher convergence accuracy compared to the standard CFOA. These test functions systematically verify the comprehensive optimization ability of the algorithm in a multi-dimensional complex environment by simulating scenarios such as ideal conditions, variable coupling, non-linear correlation, and noise interference, providing a key benchmark basis for algorithm improvement.

[0040] Step S4: Based on the improved fishing algorithm in S3, continuously optimize to determine the best learning rate and penalty factor of the LSTM model, and obtain the ECFOA-LSTM network model. The learning rate and penalty factor optimized by ECFOA can more efficiently determine the best parameter combination of the ECFOA-LSTM model in the dataset compared to manual parameter tuning. Moreover, compared to the standard fishing algorithm, the parameters optimized by the improved fishing algorithm of the present invention can more accurately capture the mapping relationship between the input and output.

[0041] Step S5: Based on the IMF component set obtained in S2, obtain the wind power prediction value. The specific steps are as follows: S51: Use the first 70% of the IMF component set as the training set and input it into the ECFOA-LSTM network model to train the model; S52: Use the remaining 30% of the IMF components in the IMF component set as the prediction set and input it into the trained ECFOA-LSTM network model to obtain the predicted values corresponding to the IMF components; S53: Add up the predicted values corresponding to all IMF components in the test set, and the sum is the wind power prediction value.

[0042] The present invention improves the frost optimization algorithm and combines it with variational mode decomposition, which can more accurately and efficiently perform feature extraction and decomposition preprocessing on the original data, effectively removing noise interference. In addition, the present invention also improves the fishing algorithm. The improved fishing algorithm can deeply optimize the hyperparameters of the LSTM model, making the network structure of the LSTM model more reasonable and the learning ability stronger, improving the adaptability and processing ability of the model to complex data, and significantly improving the prediction accuracy and stability. In tasks such as time series prediction, it has more excellent performance and generalization ability compared to single models or traditional combined models.

[0043] Experimental verification The prediction effects of the present invention and the existing technologies such as Back Propagation Neural Network (BPNN), Support Vector Machine (SVM), Long Short - Term Memory (LSTM), Catch Fish Optimization Algorithm - Long Short - Term Memory (CFOA - LSTM), and Robust Ice - cold Metaheuristic Algorithm - Variational Mode Decomposition - Catch Fish Optimization Algorithm - Long Short - Term Memory (RIME - VMD - CFOA - LSTM) are as Figure 2 shown. The comparison results are shown in Table 1. By comparing the single LSTM and CFOA - LSTM, it can be found that the three indicators of the root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) of the present invention are all smaller. Therefore, the model of the present invention performs better than other models. Taking MAPE as an example, compared with BPNN, the MAPE of the present invention is reduced by 43.37%; compared with SVM, the MAPE of the present invention is reduced by 31.58%; compared with LSTM, the MAPE of the present invention is reduced by 3.92%; compared with CFOA - LSTM, the MAPE of the present invention is reduced by 1.43%, and compared with RIME - VMD - CFOA - LSTM, it is reduced by 0.96%.

[0044] Table 1 Comparison of model indicators

[0045] The ERIME of the present invention and the RIME of the existing technology are compared on the CEC2005 (Congress on Evolutionary Computation 2005 Benchmark Functions) test function set, and the results are shown in Figure 3. It can be seen from Figure 3 that the convergence speed of the present invention is faster and the optimization accuracy is higher. The ECFOA of the present invention and the CFOA of the existing technology are also compared on the CEC2005 test function set, and the results are as Figure 4 shown. According to Figure 4It can be seen that the convergence rate of the present invention is faster and the optimization accuracy is higher.

[0046] Each embodiment in this specification is described in a related manner. For the same or similar parts among the embodiments, reference can be made to each other. Each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and reference can be made to the corresponding part of the method embodiment for the relevant content.

[0047] The above description is only for the preferred embodiments of the present invention and is not intended to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.

Claims

1. A short-term wind power prediction method, characterized in that the steps include: Step S1, collecting historical wind power data and preprocessing the data; Step S2, improving the frost and ice optimization algorithm to extract multi-frequency domain characteristics of wind power; Step S3, constructing an LSTM model and improving the fishing algorithm; Step S4, determining the optimal learning rate and penalty factor of the LSTM model based on S3, and obtaining the ECFOA-LSTM network model; Step S5, obtaining a wind power forecast value based on the IMF component set obtained in S2.

2. The short-term wind power prediction method according to claim 1, characterized in that, The specific steps of S2 are as follows: S21. Improve the frost ice optimization algorithm by changing the cosine transform of the frost ice factor to a sine-cosine dual transform, as follows: ； Among them, rand represents a random number between 0 and 1 that follows a normal distribution, Riter represents the current iteration number of the frost optimization algorithm, Max _ Riter represents the maximum iteration number of the frost optimization algorithm, W represents the algorithm parameter, and its value is 5, RimeFactor represents the frost factor, which determines the performance of the frost optimization algorithm; S22, the penalty coefficient and the number of modes in the variational mode decomposition are optimized by ERIME, the formula is as follows: ； Among them, E represents the determination coefficient of the update method, Rimepop represents the updated penalty coefficient and mode combination, best_rime represents the current best penalty coefficient and mode combination with the smallest sample entropy, represents the upper limit of the penalty coefficient and the modal number, represents the lower limit of the penalty coefficient and the modal number, rand and Represents a random number between 0 and 1 that follows a normal distribution; S23. Taking the minimum sample entropy as the objective function, perform variational mode decomposition on the wind power data in the training set and the test set to obtain the set of intrinsic mode function components of the wind power. Each IMF component characterizes the local features in different frequency domains, and continuously optimize the mode number K and the penalty coefficient through the minimum sample entropy , to find the best IMF components. The smaller the sample entropy, the better the IMF components. The specific formula is as follows: ； where SE represents sample entropy, K represents the number of modes, represents the penalty coefficient, and arg min represents the minimization operation, represents the index of the sample entropy.

3. A short-term wind power prediction method according to claim 1, characterized in that, The specific steps of the S3 improved fishing algorithm are as follows: S31, initialize the population size, the formula is as follows: ； in, Indicates The group learning rate and penalty factor are dimensional search space, Represents the learning rate and penalty factor in The upper bound of the dimensional search space, Represents the learning rate and penalty factor in the first The lower bound of the dimensional search space, rand represents a random number between 0 and 1 that follows a normal distribution; S32, the exploration phase is divided into independent search mode and group fishing mode, and the capture rate is calculated as follows: ； Among them, represents the capture rate, EFs represents the current evaluation quantity, and MaxEFs represents the maximum evaluation quantity; S33, using the differential idea to improve the independent search mode of the fishing algorithm, update the learning rate and penalty factor of the current position, the specific formula is as follows: ； ； In the formula, Exp is the empirical value parameter, and its value range is -1~1. Indicates The worst fitness value after the position update, Indicates The best fitness value after the second position change is the minimum root mean square error RMSE. represents the fitness value of the learning rate and penalty factor, Indicates The position after iteration The fitness value of the corresponding learning rate and penalty factor, Indicates the exploration range, represents the Euclidean distance between the learning rate and the penalty factor, EFs represents the current number of evaluations, and MaxEFs represents the maximum number of evaluations. represents the number of iterations of learning rate and penalty factor, Indicates The first iteration The group learning rate and penalty factor are The position in the dimensional search space, Indicates The first iteration The group learning rate and penalty factor are The position in the dimensional search space, Indicates The position after iteration The learning rate and penalty factor at The position in the dimensional search space, is a random number between 0 and 1 that follows a normal distribution; s represents a d-dimensional random unit vector, which is used to determine the main direction and distance of travel; Represents a random number from 0 to 2 that follows a normal distribution. represents a random number from 1 to 3 that follows a normal distribution, and rand represents a random number from 0 to 1 that follows a normal distribution; S34. In the simulated group fishing mode, select 3 to 4 groups of learning rates and penalty factors in sequence to form a subgroup , the subgroup obtains a discrete Laplace operator through discrete Laplace transform. The specific formula is as follows: ； ； Wherein, represents the target point surrounded by the subgroup ; represents the -th iteration of the subgroup in the -th group of learning rate and penalty factor in the -dimensional search space; represents the -th iteration of the subgroup in the -th group of learning rate and penalty factor in the -dimensional search space; represents the speed of the fisherman approaching the center, with a value range of 0 to 1, is the offset of the movement, with a value range of -1 to 1, and gradually decreases with the increase of EFs; EFs represents the current number of evaluations, MaxEFs represents the maximum number of evaluations, laplace represents the discrete Laplace operator, and A represents the matrix for randomly arranging the population, represents the -th iteration of the subgroup in the position of the learning rate and penalty factor, and mean represents the mean of the parameters, that is, the center position of the group; S35, in the simulated large-scale mining stage of the fishing algorithm, an improved diagonal form is used for updating, and the formula is as follows: ； ； In the formula, represents the variance, EFs represents the current number of evaluations, and MaxEFs represents the maximum number of evaluations. Indicates After iteration The location of the group learning rate and penalty factor, is the global optimal position, GD is the Gaussian distribution function, Represents a random number from 1 to 3 that follows a normal distribution, used to distribute the learning rate and penalty factor in three ranges; Represents the average matrix of each dimension of the learning rate and penalty factor center, diag(A) means taking the value on the diagonal after randomly permuting the population as the update condition for this round of iteration, mean represents the mean of the parameter, and rand represents a random number from 0 to 1 that follows a normal distribution.

4. A short-term wind power prediction method according to claim 1, characterized in that The specific steps of S5 are as follows: S51, input the first 70% of the IMF component set as a training set into the ECFOA-LSTM network model to train the model; S52, inputting the remaining 30% IMF components of the IMF component set into the trained ECFOA-LSTM network model as a prediction set to obtain the prediction values ​​corresponding to the IMF components; S53, adding up the predicted values ​​corresponding to all IMF components of the test set, and the sum is the wind power prediction value.

Images