EMD-Enhanced Hybrid Neural Architecture for Short-Term Wind Speed Forecasting

Accurate wind speed forecasting plays a crucial role in wind energy applications and meteorological studies. However, the inherent non-stationarity and complexity of wind speed data often limit the predictive performance of conventional methods. To address this challenge, this study introduces a EMD hybrid forecasting framework. The proposed methodology first applies Empirical Mode Decomposition (EMD) to disassemble raw wind speed signals into multiple intrinsic mode functions (IMFs), thereby mitigating data non-stationarity and noise effects. Subsequently, sample entropy analysis quantifies the complexity of each IMF, followed by a clustering-based separation into high-and low-complexity subsets. For prediction, Long Short-Term Memory (LSTM) networks handle the high-complexity components, while Autoregressive Integrated Moving Average (ARIMA) models process the low-complexity ones. The final wind speed forecast is obtained by aggregating predictions from all components.Experimental validation using a 1,500-day wind speed dataset demonstrates the superiority of the hybrid model, achieving an RMSE of 3.31, MAE of 2.65, and MAPE of 28.97— significantly outperforming standalone models. This work presents a computationally efficient and practically viable solution for wind speed prediction.