The intermittency and volatility of wind speed changes pose great challenges to the accurate prediction of wind power. Fully exploring the inherent laws of key factors such as wind power and wind speed is an effective way to improve the accuracy of wind power prediction. A method for ultra-short-term wind power prediction is proposed, which incorporates a temporal pattern attention (TPA) mechanism into a multi-layer stacked bidirectional long short-term memory network. Firstly, outlier detection for the wind power dataset is performed using a density-based noisy spatial clustering method (DBSCAN) and a linear regression algorithm, followed by data reconstruction of outlier points using k-nearest neighbor (KNN) interpolation. Next, the intrinsic correlations between wind power and various meteorological features are comprehensively considered, and the TPA mechanism is introduced into the MBLSTM network to properly allocate time step weights, capturing the underlying logical patterns of the wind power time series. Finally, the effectiveness of the proposed method is verified through experimental simulation data analysis. Results show that this method can fully explore the relationship between wind power and wind speed influencing factors, thereby improving its prediction accuracy.