In the context of energy transition and sustainable development, wind power forecasting technology is crucial for enhancing system dispatch flexibility and economic efficiency and maximizing wind energy utilization. While sensor popularity and big data and AI advancements have enhanced wind power forecasting accuracy, wind power’s stochastic and intermittent nature still challenges forecasting precision. Therefore, in this study, we propose a novel wind power forecasting model integrating a deep learning network with directed attention mechanisms. The model utilizes data obtained from the five-element meteorological sensor, the FT-WQX5 sensor, for initial input. Principal component analysis is employed for data preprocessing, the directed focused attention mechanism is introduced to enhance focus on key information, and the enhanced dynamic strategy-based pied kingfisher optimizer (EDS-PKO) is utilized for parameter optimization. The model integrates long short-term memory networks to capture temporal features. Results demonstrate that under normal weather conditions, the proposed model achieves root mean square error (RMSE) below 0.8, R-square (R2) above 90%, and mean absolute percentage error (MAPE) below 7%. Under adverse conditions, the model optimized with the improved EDS-PKO algorithm shows approximately 10% improvement in R2 and around 20% reduction in RMSE compared with other comparative models, with MAPE as low as 4.55%. This research provides a new technological approach for wind power forecasting, contributing to efficient wind energy utilization and stable grid operation.