Wind speed forecasting using dual-attention patch-based multiscale transformer

Accurate and reliable wind speed forecasting is essential for the sustainable development, efficient operation, and long-term growth of the wind power industry. This paper proposes a novel Patch-based Multi-scale Transformer model that captures dependencies in both temporal and multivariate dimensions of wind speed data for enhanced prediction. The model employs a scalable, serialized architecture composed of stacked Multi-Scale blocks. Each block first divides the input data into patches of varying sizes, and then extracts temporal and variable dependencies through dual attention mechanisms. Specifically, the self-attention mechanism captures temporal dependencies among different wind speed information patches, while a newly developed cross-attention mechanism extracts the influence of other meteorological variables on wind speed data, with particular emphasis on cross-lagged relationships. Experiments on real-world wind speed datasets demonstrate that the proposed model significantly improves prediction performance. Compared to the most recent state-of-the-art models, the MAE, RMSE, and MAPE evaluation metrics are reduced by an average of 25.90%, 10.04%, and 8.61%, respectively. Additionally, the impact of input length on the results is systematically analyzed. Finally, ablation experiments confirm the contribution of each key component in the model design, validating the effectiveness of the proposed architecture.