A novel hybrid deep learning model for multi-step wind speed forecasting considering pairwise dependencies among multiple atmospheric variables

Abstract

The reliable wind speed forecasting is critical for wind farms as it enables them to improve cost-effectiveness and optimize energy efficiency. In this study, a novel hybrid deep learning model is proposed for multi-step wind speed forecasting, which simultaneously captures pairwise dependencies and temporal features of multiple atmosphere variables. Initially, to extract intrinsic mode functions (IMFs) of nonstationary atmospheric variables, the Variational Mode Decomposition (VMD) is used. The decomposed atmospheric variables are then grouped according to similarity of mode and the pairwise dependencies are captured using the Graph Neural Network (GNN). Subsequently, the Temporal Convolutional Network (TCN) is applied for wind speed forecasting. The proposed model notably outperforms existing models with a Mean Absolute Error (MAE) of less than 0.1 m/s in one-hour ahead forecasting (1-h). Additionally, it improves wind speed forecasting accuracy by 79.22% at 1-h, 79.13% at 2-h, 80.00% at 3-h, 78.70% at 6-h, and 58.41% at 12-h, respectively, compared to the frequently used LSTM technique for wind speed forecasting. The outcomes of this study indicate that this novel approach to wind speed forecasting holds substantial potential for real-world wind farm operational contexts.