Decomposition-Driven Transformer for Accurate and Trend-Aware Wind Power Prediction

Wind power (WP) is increasingly integrated into power systems, making accurate short-term forecasting critical for economic and secure grid operation. While Transformer-Based models have reduced point-wise numerical error, they often fail to preserve trend consistency - a property more relevant for dispatch decisions than small numerical deviations. In this paper, we propose decomposition-transformer (DTrans), a trend-aware transformer variant that (1) embeds a learnable hierarchical decomposition layer in each encoder / decoder block to separate trend and seasonal components; (2) adopts a value-delta dual-output paradigm to predict both absolute power and step-wise increments; and (3) uses an adaptive multi-objective loss to jointly optimize numerical accuracy and directional reliability. Empirical studies on benchmark wind-power datasets show that DTrans significantly outperforms mainstream baselines in MAE, RMSE and R² and achieves a 38.27% relative improvement in trend-consistency metrics, thereby offering a theoretically grounded and practically deployable solution to this enduring problem in renewable-energy forecasting.