Nowcasting of Wind Energy with LSTM, BiLSTM, and Auto-Regressive Techniques

Accurate prediction of wind energy potential is crucial for optimising renewable energy investments and ensuring efficient power generation. This study addresses the challenge of wind energy forecasting using Long Short-Term Memory (LSTM), Bidirectional LSTM (BiLSTM), and Auto-Regressive (AR) models on two real-world wind energy datasets with hourly and 15-minute resolutions. Model performance is evaluated using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and the coefficient of determination $\left(R^2\right)$. Experimental results show that for Dataset-1, LSTM achieves superior long-term forecasting accuracy with an $R^2$ of 0.9749 for one-week prediction and maintains robustness for six-month forecasts $(R^2=0.9576)$, outperforming both Bil-STM and AR models. BiLSTM demonstrates marginally better short-term performance, achieving the highest $R^2$ of 0.8856 for 24-hour prediction. For Dataset-2, AR exhibits strong short-term performance with an $R^2$ of 0.9780 and MAPE of 2.17% for 24-hour forecasting, while LSTM and BiLSTM dominate mediumand long-term horizons, consistently achieving $R^2$ values above 0.99. These findings highlight the trade-offs between model complexity, accuracy, and computational efficiency, providing practical insights for selecting appropriate forecasting models across different wind energy prediction horizons.