MDHNet: Fusing global trends and local fluctuations via multi-scale decomposition for ultra-short-term wind power forecasting

Accurate ultra-short-term wind power forecasting is essential for real-time dispatch and ensuring the reliability of modern power grids. However, the predictive accuracies of forecasting methods are often compromised by the strong coupling of global trends and local fluctuations in wind power time series data. In this work, we propose a Multi-Scale Decomposition Hybrid Network (MDHNet), which employs multi-scale 1D convolutions to decouple the time series into distinct local seasonal components and global trend components. Unlike traditional fixed or manually designed decomposition schemes, the proposed multi-scale convolutional smoothing module is fully learnable and optimized end-to-end with the forecasting objective. The seasonal module identifies dominant periods via the Fourier transform, and then reshapes the sequence into a 2D tensor. Subsequently, 2D convolutions are applied to capture the complex inter-period and intra-period variations in the sequence. Concurrently with the seasonal module, the trend module extracts global long-term patterns using linear layers. These two modules are trained within independent feature spaces, and their representations are subsequently fused to yield a joint prediction. Optimized under a unified forecasting objective, the proposed MDHNet effectively integrates long-term dependencies with short-term variability. Compared to the strongest baseline, our approach demonstrates consistently superior performance, achieving average reductions of 20.73% and 21.31% in MAE and RMSE, respectively, on the Nordic dataset, while maintaining highly competitive results on the GEFCom2014 benchmark.