Collaborative wind speed forecasting based on multi-strategy MVMD and RestrictedCrossformer

Accurate and reliable wind speed forecasting is critical for optimizing wind farm operations and energy network planning. This paper proposes a hybrid deep learning model that integrates Multivariate Variational Mode Decomposition (MVMD) with a novel RestrictedCrossformer architecture. Initially, MVMD decomposes wind speed sequences from multiple observation stations with similar characteristics into several Intrinsic Mode Functions (IMFs) and residual components. Subsequently, two distinct strategies tailored to residual characteristics are implemented within the RestrictedCrossformer to jointly forecast IMFs and residual sequences. The proposed RestrictedCrossformer is designed to address the computational redundancy of the original Crossformer in modeling dependencies among IMF components through a restricted attention mechanism and the introduction of a novel Grouped Spatial Module (GSM).The GSM efficiently captures the relationships between the specified type of IMFs across different stations in a lightweight manner. Finally, predictions of all sub-components are aggregated to generate the final wind speed forecasts. Extensive experiments on two datasets demonstrate MAE reductions over 17.5 % and 55.4 %, and RMSE reductions over 17.1 % and 55.0 % compared to other MVMD-based benchmarks.