A novel spatiotemporal fusion wind speed prediction method under wind farm control center: BiTCN–transformer–cross-attention.

This article presents a hybrid deep learning model for multivariate wind speed prediction across multiple wind farms, aiming to improve forecasting accuracy and reduce operational costs at wind power control centers. The model integrates a Bi-directional Temporal Convolutional Network (BiTCN) and a Transformer with a cross-attention mechanism (BTTCA) to extract and fuse spatiotemporal features from wind speed and related meteorological variables, which are first decomposed using Multivariate Variational Mode Decomposition (MVMD). Transfer learning is applied to adapt pre-trained models from four representative wind farms to others within the control area, while the Multi-Objective Pelican Optimization Algorithm (MOPOA) optimizes the weighted fusion of predictions from these models. Experimental results using data from Hebei Province demonstrate that the proposed multivariate, hybrid approach outperforms existing models in accuracy and efficiency, with MOPOA providing superior optimization compared to other algorithms.