DMSPFF: A deep multi-scale perceptual fusion framework for short-term wind power forecasting

Accurate short-term wind-power forecasting supports power-system operation, but wind-power sequences combine slowly varying regimes, abrupt ramps, and non-stationary scale-dependent oscillations. Existing models often emphasize one scale or use fixed feature fusion, limiting their ability to model interactions among long-term context, recent dynamics, and time–frequency patterns. This study proposes a deep multi-scale perceptual fusion framework (DMSPFF). Parallel HiPPO-LegS, sLSTM, and CWT-CNN branches encode long-term memory, short-term dynamics, and localized time–frequency structure, respectively. Transformer-based attention refines interactions between long-term and time–frequency representations before their integration with short-term features. Experiments on a German meteorological and wind-power dataset evaluate 10-, 30-, 60-, and 120-min horizons against recurrent, tree-based, kernel-based, and representative recent Transformer models. At 120 min, DMSPFF achieves R 2 = 0.88 and MAE = 3.55 kW, versus the best baseline R 2 of 0.79 and the best MAE baseline of 4.80 kW, reducing MAE by 26.0%. Wind-speed-stratified evaluation supports within-dataset robustness; two-sided Diebold–Mariano tests indicate significantly lower squared-error loss in 45 of 48 comparisons at the 5% level. These results support coordinated multi-scale representation learning for longer-horizon wind-power forecasting and operational decision-making.