Multi-step wind speed forecasting based on multi-task learning using mixed granularity collaborative network

Accurate multi-step wind speed forecasting is crucial for wind farm scheduling and energy management. However, most existing studies rely on iterative multi-step forecasting at single temporal granularity, which makes it difficult to simultaneously capture short-term high-frequency fluctuations and medium-to long-term trends in wind speed. To address this issue, this paper proposes Mixed Granularity Collaborative Network (MGCNet). First, MGCNet formulates wind speed forecasting at different temporal granularities as parallel sub-tasks and jointly models short-term high-frequency components and long-term trend information within Multi-Task Learning (MTL) framework. To mitigate the scale mismatch and semantic gap between high-frequency meteorological factors and multi-granularity forecasting targets, feature alignment mechanism is designed. Building on this, MGCNet employs stacked cross-attention modules to enable dynamic information interaction among forecasting channels of different granularities, allowing tasks at different scales to share critical patterns while preserving their distinctive representations. Furthermore, to optimize the multi-task training process, this paper introduces dynamic uncertainty-based weighting strategy to balance the learning difficulty and loss contributions of the forecasting tasks at each granularity. Experimental results on wind farm dataset demonstrate that MGCNet outperforms single-granularity forecasting models and multi-step forecasting methods across different forecasting horizons and multiple evaluation metrics.