Multivariate short-term wind speed prediction based on CNN-LSTM-SA models

Abstract This paper investigates the application of the CNN-LSTM-SA model for wind speed prediction. It is found that compared with the univariate input method which only inputs the historical wind speed, the multivariate input prediction model can substantially improve the prediction accuracy. In the experiments of this paper, out of the CNN model, other multi-input prediction models improve the prediction accuracy by about 40% compared to the single-input model. In addition, this paper experimentally investigates the effect of the attention mechanism on the multi-input wind speed prediction model: the attention mechanism improves the CNN model significantly, the prediction accuracy improves by about 53%, and the degree of overfitting decreases dramatically; the effect on the LSTM model is completely negative; the improvement of the prediction accuracy of the CNN-LSTM model is not significant, and the degree of its overfitting decreases by about 66.6%.