A novel two-stage deep learning wind speed forecasting method with adaptive multiple error corrections and bivariate Dirichlet process mixture model

Highlights: • The proposed model can extract unpredictable components from raw wind speed. • Two models are built on predictable and unpredictable components respectively. • An adaptive multiple error corrections method is developed. • The bivariate Dirichlet process mixture model can produce heteroscedastic results. -- Abstract: Wind power is promising renewable energy. Wind speed forecasting is essential for wind energy integration and application. The wind speed series can be divided into two components, including predictable and unpredictable components. In this study, a novel two-stage forecasting model is proposed for dealing with those two components separately. In the first stage, a novel model consisting of wavelet packet decomposition, convolutional neural network and adaptive multiple error corrections is proposed to forecast the predictable components. The adaptive multiple error corrections method can eliminate the predictable components in forecasting residuals of the wavelet convolutional neural network model, which is unique. In the second stage, bivariate Dirichlet mixture model is proposed to model heteroscedasticity of the unpredictable residuals with the non-parametric distribution. Several important hyper-parameters are selected by blocked cross validation. Three actual wind speed series are utilized to verify the effectiveness of the proposed model. The results show that the proposed model outperforms the benchmark models.