A Hybrid Wind Speed Prediction Model Based on Signal Decomposition and Deep 1DCNN

Abstract

AbstractWind speed prediction is a typical time series prediction and is of great importance in power generation. In order to deal with those problems of heavy resource consumption and complex hyperparameter selection in traditional methods, we propose a multidimensional prediction method based on decomposition methods. However, using a model to fit all subseries may lead to the model’s performance degradation and error increasing, which is called “preference” error. To solve this problem, a one-dimensional CNN (1DCNN) is used to capture the relationships between subseries. As to better explore this problem and enhance the stability of the CNN model, the generative adversarial network (GAN) method is tried to generate and generalize this “preference” error and expand training samples for 1DCNN. This paper combines multiple methods including the decomposition method, RNN model, CNN model, and GAN method in order, and chooses the best combination in different datasets. The experiments on two real-world wind datasets demonstrate that this method can achieve excellent performance in wind speed prediction with the help of combining the above methods.KeywordsWind speed predictionSignal decompositionLSTMCNNWGAN