This paper introduces a sophisticated method for ultra-short-term wind power prediction by integrating time series decomposition interaction mechanisms with a Dual Exponential Moving Average (EMA) Attention Module. Initially, the input data are divided into trend and seasonal components, each of which is processed by dedicated modules. Linear transformations are applied to the trend component, while the seasonal component is handled by a Seasonal Processing Module that employs the Dual EMA Attention mechanism to capture both extended trends and detailed short-term fluctuations. Experimental results indicate that this method significantly enhances the accuracy and stability of the forecast, surpassing the baseline models, particularly in addressing the inherent variability of the wind power data.