Enhanced Wind Power Prediction Using LSTM with Self-Attention Mechanism: A Hybrid Approach for Improved Accuracy and Reliability

Accurate wind power forecasting is paramount for stable grid integration and efficient energy management. However, existing methodologies are often constrained by the inherent variability of meteorological data and complex temporal dependencies. We propose a novel model based on Long Short-Term Memory (LSTM) networks, enhanced by a self-attention mechanism, to effectively capture long-range dependencies and dynamically weight relevant temporal features. This approach employs a dual-LSTM architecture to process serialized wind power data: the initial layer extracts hierarchical temporal features, while the subsequent layer refines them through an attention-based feature recalibration. The self-attention mechanism explicitly models interactions between distant time steps, enabling the model to focus on critical historical patterns and mitigate noise interference. Experimental results demonstrate that our method outperforms baseline XGBoost and Random Forest models in terms of both prediction accuracy and stability, exhibiting particular strength in scenarios characterized by highly volatile wind speeds. Furthermore, visualization of attention weights enhances the model's interpretability, intuitively revealing key meteorological factors influencing power generation. This study contributes a balanced architecture regarding complexity and performance to the renewable energy forecasting domain, offering practical value for grid operators in obtaining reliable short-term wind power estimations. The proposed solution has been validated using operational data from wind farms with diverse climatic characteristics, affirming its feasibility for practical deployment.