Short-Term Wind Speed Hybrid Prediction Model Based On Multiple Decomposition

Wind power generation,as a highly promising new energy generation technology,is gradually developing into a mainstream energy source worldwide.However,the inherent stochastic variability of wind poses serious challenges to the safe operation of the power grid and the reliability of power supply.Improving the accuracy of wind speed prediction is one effective way to address this issue.However,most wind speed prediction models cannot well capture the inherent regularity of wind speed data.In order to fully utilize deep features,reduce computational costs,and improve prediction accuracy,this paper studies wind speed point prediction and interval prediction,and combines decomposition algorithms,variable selection methods,artificial intelligence models,and statistics based on wind speed prediction information to propose a new hybrid prediction model,providing a model reference for other similar time series problems.(1)Multiple decomposition of wind speed sequence.The original wind speed sequence is decomposed into multiple sub-sequences through quadratic decomposition and sample entropy analysis,further extracting the frequencydomain features of wind speed fluctuations.After EMD decomposition,the unpredictability of each wind speed sub-sequence is estimated by sample entropy,and the VMD is used to further decompose the sub-sequence with the minimum unpredictability to obtain secondary sub-sequences and further extract wind speed fluctuation features to improve the prediction ability of the model.(2)Construction of model inputs and outputs.Considering the issue of time delay,the sub-components obtained from the decomposition are reconstructed into a phase space matrix through Phase Space Reconstruction(PSR).The onedimensional wind speed data is mapped to a high-dimensional space for prediction,further reducing the difficulty of prediction.(3)Sub-sequence prediction.The Nbeats algorithm is used to predict the sub-sequence.This neural network structure is based on forward and backward residual connections and a deep fully-connected layer stack.The internal residual blocks can solve various problems,provide interpretable outputs,and use as little prior knowledge as possible,i.e.significantly reduce the amount of training set data and shorten training time.(4)The final prediction result is derived by accumulating the predicted values of all sub-components.(5)Modeling of sample data for four seasons and comparison with multiple comparison models using an improved evaluation criterion to test the predictive performance of the model.The proposed predictive method fully excavates the potential information of wind speed and has strong stability and high prediction accuracy,verifying the superiority of the proposed hybrid method.