Short-term wind speed prediction with adaptive signal processing based hybrid statistical models

Abstract

The inherent nonlinearity, intermittency, and chaotic nature of wind speed make accurate forecasting challenging. Traditional approaches like standalone time series models and frequency domain analysis struggle to capture these complex characteristics effectively. In light of this, the present study utilizes three self-adaptive signal processing methods, namely empirical mode decomposition (EMD), ensemble empirical mode decomposition (EEMD), and variational mode decomposition (VMD) and combines with ARIMA or window-sliding ARIMA (WSARIMA) to develop six hybrid models, namely EMD–ARIMA, EEMD–ARIMA, VMD–ARIMA, EMD–WSARIMA, EEMD–WSARIMA, and VMD–WSARIMA. To illustrate the efficacy of the proposed hybrid models in daily wind speed prediction, four study sites from India with different climates are considered. Based on the analysis of 7 years (08-2015–03-2023) of wind speed data, it is found that: (i) the extracted components of VMD overcome the limitations of EMD and EEMD methods; (ii) the combination of VMD and WSARIMA outperforms any other comparative model, such as ARIMA, WSARIMA, EMD–ARIMA, EEMD–ARIMA, VMD–ARIMA, EMD–WSARIMA, or EEMD–WSARIMA; the VMD–WSARIMA model reduces RMSE by 70–80% compared to the conventional ARIMA model; (iii) finally, as a part of post-processing, the residual analysis of the best fit VMD–WSARIMA model shows desirable characteristics. Therefore, the present study strongly recommends to consider adaptive decomposition based hybrid models in wind speed forecasting at shorter time horizon.