Windformer: A novel spatiotemporal wind speed forecasting model based on Transformer for wind energy management

This paper proposes a novel spatiotemporal wind speed prediction model, named Windformer, which introduces targeted improvements based on the classical Transformer architecture to better capture the complex spatiotemporal dependencies inherent in wind speed data. Traditional wind speed prediction methods and the classical Transformer model have certain limitations when dealing with long-term sequence dependencies and multi-scale spatial features. To address these issues, Windformer incorporates a Multi-Scale Spatiotemporal Adaptive Attention Mechanism (MSA-Attention), which adaptively extracts wind speed features across different temporal scales and spatial ranges. In addition, Windformer employs a Wind Field-Aware Positional Encoding, enhancing the model’s ability to represent dynamic wind field scenarios, thereby enabling more precise capturing of wind speed variations. Experimental results demonstrate that Windformer significantly outperforms traditional methods and the original Transformer architecture in terms of prediction accuracy and stability across multiple real-world wind speed datasets, showcasing its great potential in wind energy prediction and optimized management. This study provides a new approach and perspective for intelligent wind energy management and the efficient utilization of renewable energy resources.