Enhancing Wind Power Forecasting in Power Grids With Dual Attention-based LSTM Mechanism

Abstract

Power grid operators must deal with sustainable energy integration challenges because wind power supply and demand patterns show natural variability. The precise prediction of wind power supplies is critical to grid efficiency because it controls how renewable energy sources join the network system. Using long short-term memory (LSTM) networks within deep learning techniques shows high effectiveness for predicting short-term wind power forecasting. The paper presents a sophisticated dual-attention LSTM system that enhances wind power prediction accuracy. The model uses LSTM networks to detect temporal patterns and dual attention mechanisms to choose essential features that lead to improved wind power predictions during times of variable conditions. The proposed dual-attention LSTM model outperforms the current models of forecasting regarding one-step and multi-step wind power prediction. Trained and tested on Texas Wind Turbine dataset released publicly, the model produced an RMSE of 140.79 and 155.23 in single-step and multi-step forecasting, respectively, and consistently lower values of both MAE and MAPE, as compared to all baseline models. These findings reveal the value of the dual attention mechanism to manage the variable wind conditions and future directions are expected to be domain adaptation and transfer learning to enable real-time deployment, multi-site operation.