A medium- and long-term wind speed prediction method and system

Abstract

The present invention is applicable to the field of wind speed prediction technology and provides a medium- and long-term wind speed prediction method and system, comprising the following steps: collecting wind speed data and constructing an input-output dataset; using an artemisinin optimization algorithm to find the optimal number of intrinsic mode components K and penalty coefficient α in VMD to obtain IMF components; constructing an LSTM network model, optimizing the number of hidden units, maximum training cycle, and initial learning rate of the LSTM network model as the objectives of the mirage optimization algorithm, retaining the optimal parameters of the LSTM network model, and obtaining a FATA-LSTM network model; training, testing, and validating the FATA-LSTM network model using the IMF components to generate a wind speed prediction model; and outputting prediction results. The present invention effectively extracts multi-scale features and reduces noise interference, improving generalization capability, and achieving higher prediction accuracy in complex wind speed environments.

Description

Technical Field

[0001] The present invention belongs to the technical field of wind speed prediction, and in particular relates to a medium- and long-term wind speed prediction method and system. Background Art

[0002] Wind energy plays a vital role in addressing carbon emissions and energy shortages. However, its significant non-stationary and intermittent characteristics pose challenges to the smooth operation of power grids, power generation efficiency, and power quality. Accurate wind speed prediction technology is crucial for ensuring stable power system operation and optimizing economic dispatch. It can help wind farms develop scientific and reasonable control strategies to improve wind energy utilization efficiency.

[0003] Current wind speed prediction methods include physical models (such as numerical weather forecasting), statistical models (such as ARIMA) and intelligent models based on machine learning. In recent years, hybrid models that combine signal decomposition and deep learning have attracted widespread attention due to their high accuracy.

[0004] However, existing medium- and long-term wind power forecasting technologies have bottlenecks: the variational mode decomposition (VMD) method lacks adaptive capabilities and is prone to excessive or incomplete filtering in signal sources with low signal-to-noise ratios and frequency domain aliasing. At the same time, the estimation of the number of signal sources is often not accurate enough. When performing signal source decomposition, problems such as incorrect signal source filtering, inaccurate estimation of the number of signal sources, and excessive time-frequency domain distribution often occur; the fitting ability of the LSTM model is highly dependent on the setting of hyperparameters, and traditional hyperparameter tuning methods are inefficient and inaccurate; real-time monitoring and prediction methods based on sensor data are costly, limited by equipment accuracy and reliability, have low economic efficiency, and have limited promotion prospects.

[0005] In summary, the existing technologies in medium- and long-term wind power generation forecasting mainly have problems such as low prediction accuracy and poor economic efficiency, which makes it difficult to effectively support the further development of wind power generation. Summary of the Invention

[0006] The purpose of the embodiments of the present invention is to provide a medium- to long-term wind speed prediction method, aiming to solve the problems raised in the above background technology.

[0007] The embodiment of the present invention is implemented as follows: a medium- and long-term wind speed prediction method includes the following steps:

[0008] Collect wind speed data and construct input-output data sets;

[0009] The artemisinin optimization algorithm is used to find the optimal number of intrinsic mode components K and penalty coefficient α in VMD to obtain the IMF components;

[0010] Build an LSTM network model, use the number of hidden units, maximum training cycle, and initial learning rate of the LSTM network model as the objectives of the mirage optimization algorithm to optimize, retain the optimal parameters of the LSTM network model, and obtain the FATA-LSTM network model;

[0011] The FATA-LSTM network model is trained, tested, and validated using the IMF component to generate a wind speed prediction model;

[0012] Output the prediction results.

[0013] Another object of an embodiment of the present invention is to provide a medium- and long-term wind speed prediction system for implementing the above-mentioned medium- and long-term wind speed prediction method, comprising:

[0014] The acquisition module is used to collect wind speed data and construct input-output data sets;

[0015] IMF component acquisition module, used to find the optimal intrinsic mode component number K and penalty coefficient α in VMD using artemisinin optimization algorithm to obtain IMF components;

[0016] The FATA-LSTM network model construction module is used to construct the LSTM network model. The number of hidden units, maximum training cycle, and initial learning rate of the LSTM network model are used as the objectives of the mirage optimization algorithm to optimize the model. The optimal parameters of the LSTM network model are retained to obtain the FATA-LSTM network model.

[0017] The wind speed prediction model generation module is used to train, test and verify the FATA-LSTM network model using the IMF component to generate a wind speed prediction model;

[0018] Output module, used to output prediction results.

[0019] The present invention provides a medium- and long-term wind speed forecasting method that uses the FATA algorithm to optimize the number of hidden units, maximum training cycle, and initial learning rate of the LSTM network model, thereby solving the problem of difficult hyperparameter selection. By predicting the IMF component, the method can accurately output wind speed change trends in real time.

[0020] By combining the AO-VMD decomposition technology to preprocess the original wind speed signal, multi-scale features can be effectively extracted and noise interference can be reduced, thereby improving the generalization ability of the model and making the model more accurate in predicting complex wind speed environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 A flow chart of a medium- to long-term wind speed prediction method provided by an embodiment of the present invention;

[0022] Figure 2 A structural diagram of a medium- and long-term wind speed prediction system provided by an embodiment of the present invention;

[0023] Figure 3 A graph of summer wind speed data provided by an embodiment of the present invention;

[0024] Figure 4 A comparison chart of the prediction results of various models provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0025] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0026] The specific implementation of the present invention is described in detail below with reference to specific embodiments.

[0027] like Figure 1 FIG. 1 is a flow chart of a method for medium- to long-term wind speed prediction according to an embodiment of the present invention, comprising the following steps:

[0028] S1. Collect historical wind speed time series data of the target wind farm, construct input-output data sets, and prepare data for subsequent decomposition and prediction:

[0029] Get hourly wind speed records, recorded as raw signal data, with a length of T;

[0030] The sliding window method is used to construct the data set. The wind speed data for 24 consecutive hours is used as the input feature x. The wind speed value y for the next hour is predicted, and the feature matrix x and target vector y with a sample number of T-24 are generated.

[0031] S2. Use the artemisinin optimization algorithm (AO) to optimize the variational mode decomposition (VMD) to perform modal decomposition on the wind speed data and influencing factors, obtain a set of IMF components with different frequency domain characteristics, and use the artemisinin optimization algorithm to find the optimal intrinsic mode component number K and penalty coefficient α by dynamically adjusting the penalty factor α and the mode number K:

[0032] S201. Define the range of the joint optimization parameter set of VMD decomposition core parameters: wherein the range of K is 2 to 30, and the range of α is 500 to 8000, which respectively represent the preset lower limit to upper limit of the decomposition mode number and the preset lower limit to upper limit of the bandwidth constraint;

[0033] S202, using the good point set strategy to initialize the artemisinin population, can overcome the problem that the original AO based on the uniform random number initialization process cannot achieve complete coverage of the search space, thereby avoiding key problems such as poor convergence performance and difficulty in maintaining population diversity. Each artemisinin agent represents a set of VMD parameter combinations. The specific process is:

[0034] Determine the population size N=2, problem dimension dim=2 (K and α in solving the two problem VMD parameters) and the maximum number of iterations MaxFEs=20;

[0035] In the parameter range (parameter Find the smallest prime number greater than 2*dim+3 within the lower bound lb=[2,500] and upper bound ub=[30,8000] To generate the benchmark for the good point set, the initial population pop is randomly generated, which is expressed as:

[0036] ;

[0037] ;

[0038] ;

[0039] in, ; ; ;

[0040] S203, executing the comprehensive elimination phase, simulating the comprehensive elimination effect of artemisinin in the early stage of malaria treatment, and searching the VMD parameter space through the global perturbation strategy:

[0041] S2031. Use the good point set strategy to initialize the artemisinin population and calculate the fitness of each agent: , the objective function is the reconstruction error of vmd decomposition:

[0042] ;

[0043] Where, is the original wind speed signal; is the kth IMF component;

[0044] Get the current best agent best and best fitness Bestfitness, corresponding to the current optimal VMD parameter combination ;

[0045] S2032, enter the iterative loop, calculate the dynamic control parameters, and update the dynamic control parameters according to the current function evaluation times FEs and the maximum evaluation times MaxFEs and the disturbance amplitude control parameter :

[0046] ;

[0047] ;

[0048] S2033. Traverse each artemisinin agent i and dimension j (j=1 corresponds to K, j=2 corresponds to α), perform VMD decomposition on each agent position pop(i,j), calculate fitness Fitness(i), and evaluate whether the position needs to be updated;

[0049] S2034, determine and update the proxy position, if the random number , perform a global perturbation update:

[0050] ;

[0051] ;

[0052] S204, executing the local clearance phase, simulating the local clearance effect of artemisinin in a specific infected area, and optimizing VMD parameters through a local adjustment strategy:

[0053] S2041. Calculate normalized fitness:

[0054] ;

[0055] like , perform a local perturbation update:

[0056] ;

[0057] Where A is the random arrangement of population index; , rand is a random number between 0 and 1;

[0058] S205. Execute the post-consolidation phase, simulating the residual sweeping and reviewing effects of artemisinin in the final phase, and optimize VMD parameters through mutation and boundary processing mechanisms (whether the previous comprehensive elimination, local clearance, or mutation, may cause the particle position to exceed the predefined search space boundary. Therefore, this embodiment of the present invention uses a boundary processing mechanism to adjust all out-of-bounds dimensions back to the legal range):

[0059] S2051, complete the mutation position update:

[0060] ;

[0061] in, is the original location of the population; is the current global optimal solution;

[0062] S2052, check whether the population has crossed the boundary:

[0063] ;

[0064] If not satisfied, do the following:

[0065] Set a random number to randomly decide whether to execute the reflection strategy or the adaptive reset strategy. Set the probability of executing the reflection strategy to 30% and the adaptive reset strategy to 70%:

[0066] Reflection strategy:

[0067] If the upper bound is exceeded:

[0068] ;

[0069] If the lower bound is exceeded:

[0070] ;

[0071] To perform a clamp operation:

[0072] If the initial Far beyond the boundary, a single reflection may cause Falling outside the boundary on the other side, the clamping operation is used to ensure that the reflected value is still in the valid search space Inside, the formula is:

[0073] ;

[0074] Adaptive reset strategy:

[0075] ;

[0076] In the metaheuristic optimization algorithm, the position update operation of the population (such as movement, mutation, etc.) may cause it to go out of bounds, that is, the coordinates of the population in a certain dimension exceed the predefined upper and lower limits of the search space. Therefore, how to deal with these transboundary populations efficiently and reasonably is a key issue that must be faced;

[0077] The Artemisinin algorithm uses a relatively simple and straightforward strategy: when a population crosses a boundary in a dimension, it is forcibly reset to the position of the current global optimal solution (best) in the corresponding dimension. This brings obvious limitations: loss of diversity, premature convergence, and reduced global search capability. To overcome the shortcomings of the original AO, the present invention introduces a boundary processing method that combines reflection and adaptive reset mechanisms:

[0078] Reflection mechanism: When a population crosses the boundary, a reflection operation is performed with a low probability. The core idea is to "bounce" the out-of-bounds population back into the search space from the boundary, just like light hitting a mirror.

[0079] Adaptive reset: When a population crosses a boundary, it is reset to the position of the current global optimal solution in the corresponding dimension. This ensures that the vast majority of out-of-bounds populations can be quickly guided to the most promising area currently known, accelerating the algorithm's convergence and preventing the population from getting "lost" in the search space.

[0080] S205. Determine whether the current evaluation number exceeds the maximum evaluation number. If so, take the current global optimal position as the final output to obtain the optimized VMD parameter combination. Otherwise, return to step S202 to continue optimizing.

[0081] S3. Build an LSTM network model. Use the number of hidden units, maximum training cycle, and initial learning rate of the LSTM network model as the targets of the mirage optimization algorithm for optimization. Retain the optimal parameters of the LSTM network model and obtain the FATA-LSTM network model:

[0082] S301, the number of hidden units of the LSTM network model , maximum training cycle and the initial learning rate As the light individual in the mirage optimization algorithm, the population is initialized as follows:

[0083] Determine the population size N=10, the maximum number of evaluations MaxFEs=20, within the parameter range (i.e. parameter , upper boundary ) randomly generates the initial population:

[0084] ;

[0085] Where, Indicates the position of the i-th ray individual in the j-th dimension; and The upper and lower bounds of the two points in the j-th dimension respectively; r is a random number between (0,1); i=1, 2, ..., N;

[0086] The mirage optimization algorithm is divided into a refraction phase and a reflection phase, which are determined by a random number r1 (between (0, 1)) and the group quality factor IP, as well as a random number r2 (between (0, 1)) and the body quality factor p. If r1>IP, it is the refraction phase and enters the direct refraction mode; if r1≤IP and r2<p, it enters the indirect refraction mode; if r1≤IP and r2≥p, it enters the total internal reflection mode.

[0087] S302, the refraction simulation stage, simulates the optical refraction process of a mirage. In the initial optimization stage, individual light explores the hyperparameter space through refraction behavior, covering a wide range. As the iteration proceeds, the refraction range gradually converges. The refraction stage is divided into direct refraction mode and indirect refraction mode. The conversion between direct refraction mode and indirect refraction mode is determined by the group quality factor IP and the individual quality factor p. The specific process includes:

[0088] S3021, simulation of direct refraction stage, initialization , calculate the fitness of each light individual:

[0089] , the objective function is the mean square error of LSTM on the test set:

[0090] ;

[0091] Where, The original data test set data; Predict data for the model;

[0092] Get the current global optimal position gBest and optimal fitness gBestScore, and sort the fitness , calculate the population quality factor:

[0093] ;

[0094] Where, Expressed as the sum of the cumulative integrals of the current population; The worst points value in history; The best points value in history; is the worst integral value, is a very small positive number;

[0095] The random number r1 is compared with IP. If r1>IP, the direct refraction mode is entered and the updated position is:

[0096] ;

[0097] Where rand is a random number between (0, 1). If the new position fitness is better than the current one, update ;

[0098] S3022. Simulate the indirect refraction mode. If r1≤IP, calculate the individual quality factor p:

[0099] ;

[0100] The random number r2 is compared with p. If r2 < p, enter the indirect refraction mode, and update the position as follows:

[0101] ;

[0102] In the formula, is the value of the global optimal position in the j-th dimension, , ;

[0103] If the fitness of the new position is better than the current one, update .

[0104] S303. Simulation reflection stage: Simulate the total internal reflection process of a mirage, and adjust the LSTM hyperparameters through light reflection:

[0105] Initialize the population using the result of step S302. If r1 ≤ IP and r2 ≥ p, enter the reflection stage, and update the position as follows:

[0106] ;

[0107] ;

[0108] In the formula, is the random reference ray individual index; ; ; is the reflection coefficient;

[0109] If the fitness of the new position is better than the current one, update ;

[0110] S304. Determine whether the current evaluation count FEs exceeds the maximum evaluation count MaxFEs. If it exceeds, determine the current global optimal position as the best hyperparameter combination of the LSTM network model, otherwise return to step S302 to continue optimization.

[0111] S4. Use the optimized , and to configure the LSTM network. Divide the IMF components decomposed in step S2 into training set, validation set and test set according to a certain proportion. Use the training set to train the FATA-LSTM network model, use the test set to test the FATA-LSTM network model, and use the validation set to validate the FATA-LSTM network model to obtain the wind speed prediction model;

[0112] S5. Input the real-time collected wind speed data into the optimized prediction model and output the prediction result.

[0113] like Figure 2 FIG. 1 is a block diagram of a medium- to long-term wind speed prediction system according to an embodiment of the present invention, comprising:

[0114] The acquisition module 100 is used to collect wind speed data and construct an input-output data set;

[0115] The IMF component acquisition module 200 is used to optimize VMD by using the artemisinin algorithm to find the optimal number of intrinsic mode components K and penalty coefficient α, and obtain IMF components;

[0116] The FATA-LSTM network model construction module 300 is used to construct an LSTM network model, optimize the number of hidden units, maximum training cycle, and initial learning rate of the LSTM network model as the objectives of the mirage optimization algorithm, retain the optimal parameters of the LSTM network model, and obtain the FATA-LSTM network model;

[0117] The wind speed prediction model generation module 400 is used to train, test and verify the FATA-LSTM network model using the IMF component to generate a wind speed prediction model;

[0118] The output module 500 is used to output the prediction results.

[0119] Performance Verification:

[0120] The wind speed data of Changchun City from April 1 to June 30, 2023 is used as the verification data. The original data is as follows: Figure 3 As shown, the wind speed prediction model provided by the embodiment of the present invention and various network models such as LSTM are used for prediction, and the results are as follows Figure 4 The prediction results of each model are shown in Table 1:

[0121] Table 1

[0122]

[0123] It can be seen that the prediction model provided by the embodiment of the present invention has a root mean square error of 0.23145, a mean absolute error of 0.18501, and a mean absolute percentage error of 3.4809%, which are all better than the single LSTM, VMD-LSTM, and AO-VMD-LSTM models.

[0124] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A medium- to long-term wind speed prediction method, characterized in that: The following steps are involved: Collect wind speed data and construct input-output data sets; The artemisinin optimization algorithm is used to find the optimal number of intrinsic mode components K and penalty coefficient α in VMD to obtain the IMF components; Build an LSTM network model, use the number of hidden units, maximum training cycle, and initial learning rate of the LSTM network model as the objectives of the mirage optimization algorithm to optimize, retain the optimal parameters of the LSTM network model, and obtain the FATA-LSTM network model; The FATA-LSTM network model is trained, tested, and validated using the IMF component to generate a wind speed prediction model; Output prediction results; The step of using the artemisinin optimization algorithm to find the optimal number of intrinsic mode components K and penalty coefficient α in VMD and obtain IMF components specifically includes: Define the range of the joint optimization parameter set of VMD decomposition core parameters: where K ranges from 2 to 30, and α ranges from 500 to 8000, representing the preset lower and upper limits of the number of decomposed modes and the preset lower and upper limits of the bandwidth constraint, respectively; The artemisinin population is initialized using the good point set strategy, and each artemisinin agent represents a set of VMD parameter combinations. ; Execute the comprehensive elimination phase, simulate the comprehensive elimination effect of artemisinin in the early stage of malaria treatment, and search the VMD parameter space through the global perturbation strategy; Execute the local clearance phase to simulate the local clearance effect of artemisinin in specific infected areas and optimize VMD parameters through local adjustment strategies; The late consolidation phase was performed to simulate the residual sweeping and review effects of artemisinin in the final stage and to optimize VMD parameters through the variation and boundary processing mechanism; Determine whether the current number of evaluations exceeds the maximum number of evaluations. If so, take the current global optimal position as the final output to obtain the optimized VMD parameter combination. Otherwise, continue to search for the optimal solution. The steps of constructing an LSTM network model, optimizing the number of hidden units, maximum training cycle, and initial learning rate of the LSTM network model as the objectives of the mirage optimization algorithm, retaining the optimal parameters of the LSTM network model, and obtaining the FATA-LSTM network model specifically include: The number of hidden units, maximum training cycle and initial learning rate of the LSTM network model are used as the light individuals in the mirage optimization algorithm to initialize the population. Specifically: Determine the population size N=10, the maximum number of evaluations MaxFEs=20, within the parameter range , 、 Generate the initial population randomly: ； Where, Indicates the position of the i-th ray individual in the j-th dimension; and The upper and lower bounds of the two points in the j-th dimension respectively; r is a random number between (0,1); i=1, 2, ..., N; The refraction simulation stage simulates the optical refraction process of a mirage, which is divided into direct refraction mode and indirect refraction mode. The LSTM hyperparameters are optimized through light propagation. In the reflection simulation phase, the total internal reflection process of a mirage is simulated, and the LSTM hyperparameters are adjusted through light reflection. Determine whether the current number of evaluations FEs exceeds the maximum number of evaluations MaxFEs. If so, determine the current global optimal position as the optimal hyperparameter combination of the LSTM network model; otherwise, continue optimization.

2. The medium- and long-term wind speed prediction method according to claim 1, characterized in that: The steps of collecting wind speed data and constructing an input-output data set specifically include: Get hourly wind speed records, recorded as raw signal data, with a length of T; The sliding window method is used to construct the data set. The wind speed data for 24 consecutive hours is used as the input feature x. The wind speed value y for the next hour is predicted, and the feature matrix x and target vector y with a sample number of T-24 are generated.

3. The medium- and long-term wind speed prediction method according to claim 1, characterized in that: The artemisinin population is initialized using the good point set strategy, and each artemisinin agent represents a set of VMD parameter combinations. The steps include: Determine the population size N = 2, the problem dimension dim = 2, and the maximum number of iterations MaxFEs = 20; In the parameter range Find the smallest prime number greater than 2*dim+3 To generate the benchmark for the good point set, the initial population pop is randomly generated, which is expressed as: ； ； ； Where, ; ; .

4. The medium- and long-term wind speed prediction method according to claim 3, characterized in that: The step of performing the comprehensive elimination stage, simulating the comprehensive elimination effect in the initial stage of artemisinin treatment of malaria, and searching the VMD parameter space through the global perturbation strategy specifically includes: Initialize the artemisinin population using the good point set strategy and calculate the fitness of each agent: , the objective function is the reconstruction error of vmd decomposition: ； Where, is the original wind speed signal; is the kth IMF component; Get the current best agent best and best fitness Bestfitness, corresponding to the current optimal VMD parameter combination ; Enter the iterative loop, calculate the dynamic control parameters, and update the dynamic control parameters according to the current function evaluation times FEs and the maximum evaluation times MaxFEs and the disturbance amplitude control parameter : ； ； Traverse each artemisinin agent i and dimension j, determine and update the agent position, if the random number , perform a global perturbation update: ； 。 5. The medium- and long-term wind speed prediction method according to claim 4, characterized in that: The step of performing the local clearance stage, simulating the local clearance effect of artemisinin in a specific infected area, and optimizing the VMD parameters through the local adjustment strategy specifically includes: Calculate the normalized fitness: ； like , perform a local perturbation update: ； Where A is the random arrangement of population index; , rand is a random number between 0 and 1.

6. The medium- and long-term wind speed prediction method according to claim 5, characterized in that: The step of performing the late consolidation stage, simulating the effect of artemisinin in the final stage, and optimizing the VMD parameters through the mutation and fusion reflection and the adaptive boundary processing mechanism specifically includes: Complete the update of the mutation position: ； Where, is the original location of the population; is the current global optimal solution; Check whether the population exceeds the boundary: ； If not satisfied, perform the following operations: Set a random number to randomly determine whether to execute the reflection strategy or the adaptive reset strategy: Reflection strategy: If it exceeds the upper bound: ； If it exceeds the lower bound: ； Perform the clamping operation, and the formula is: ； Adaptive reset strategy: 。 7. The medium- and long-term wind speed prediction method according to claim 1, characterized in that: The step of simulating the refraction stage, simulating the optical refraction process of a mirage, which is divided into a direct refraction mode and an indirect refraction mode, and optimizing the LSTM hyperparameters through the light propagation specifically includes: Simulate direct refraction stage, initialize , calculate the fitness of each light individual: , the objective function is the mean square error of LSTM on the test set: ； Where, The original data test set data; Predict data for the model; Get the current global optimal position gBest and optimal fitness gBestScore, and sort the fitness , calculate the population quality factor: ； Where, Expressed as the sum of the cumulative integrals of the current population; The worst points value in history; The best points value in history; is the worst integral value; is a very small positive number; Compare the random number r1 with IP. If r1 > IP, enter the direct refraction mode, and update the position as: ； Where rand is a random number between (0, 1). If the new position fitness is better than the current one, update ; Simulate the indirect refraction mode. If r1 ≤ IP, calculate the individual mass factor p: ； Compare the random number r2 with p. If r2 < p, enter the indirect refraction mode, and update the position as: ； Where, , ; If the new position fitness is better than the current one, update .

8. The medium- and long-term wind speed prediction method according to claim 7, characterized in that: The step of simulating the reflection stage, simulating the total internal reflection process of a mirage, and adjusting the LSTM hyperparameters through the light reflection specifically includes: If r1 ≤ IP and r2 ≥ p, enter the reflection stage, and update the position as: ； ； Where, is the individual index of the random reference ray; ; ; is the reflection coefficient; If the new position fitness is better than the current one, update .

9. An AO-VMD-FATA-LSTM medium- and long-term wind speed forecasting system, used to implement the medium- and long-term wind speed forecasting method according to any one of claims 1 to 8, characterized in that: Include: The acquisition module is used to acquire wind speed data and construct an input-output data set; The IMF component acquisition module is used to use the artemisinin optimization algorithm to find the optimal number of intrinsic mode functions K and the penalty coefficient α in VMD, and obtain the IMF components; The FATA-LSTM network model construction module is used to construct an LSTM network model, optimize the number of hidden units, the maximum training period, and the initial learning rate of the LSTM network model as the objectives of the mirage optimization algorithm, retain the optimal parameters of the LSTM network model, and obtain the FATA-LSTM network model; The wind speed prediction model generation module is used to train, test, and verify the FATA-LSTM network model using the IMF components, and generate a wind speed prediction model; The output module is used to output the prediction result.

Images