Abstract: Aiming at the insufficient prediction accuracy caused by wind speed's nonlinearity, non-stationarity, and multi-source meteorological coupling, this paper proposes a dual-branch DTCN-Informer model based on sliding-window ICEEMDAN decomposition and multi-head cross-attention (MHCA) fusion. The model first uses Lasso to select key meteorological variables, then employs sliding-window ICEEMDAN to dynamically decompose the wind speed series to reduce non-stationarity. Subsequently, a dual-branch parallel architecture is constructed: the DTCN branch extracts short-term high-frequency features through dual-dilated convolution, while the Informer branch captures long-term trends based on ProbSparse attention. Finally, a multi-head cross-attention mechanism is applied to adaptively fuse local details with global patterns, enhancing the model's generalization ability. Experimental results show that compared to the baseline Informer model, the proposed model improves MAE, MSE, RMSE, and MAPE by 62.3%, 89.7%, 67.9%, and 42.6%, respectively.