Residual hybrid deep learning for high-accuracy wind forecasting in coastal and tropical climates

Accurate wind speed forecasting is essential for optimizing renewable energy systems, managing urban microclimates, and improving meteorological preparedness. This study assesses the prediction accuracy of four advanced deep learning (DL) models—Convolutional Neural Network (CNN) + Bidirectional Long Short-Term Memory (BiLSTM) + Attention, BiLSTM, CNN-Long Short-Term Memory (LSTM), and a Residual Hybrid framework—for forecasting the 10-meter wind speed range (WS10M_RANGE) across eight climatically diverse regions in Malaysia over 24 years (2001–2024). A comprehensive set of meteorological inputs was used, including solar irradiance, temperature data, humidity, pressure, wind direction, and time indicators. The Residual Hybrid model consistently outperformed CNN + BiLSTM + Attention, BiLSTM, and CNN–LSTM architectures, achieving about 57–60% reductions in root-mean-square error (RMSE) and lowering mean absolute percentage error (MAPE) from approximately 31–35% to as low as 10.6%. Taylor diagram analysis further confirmed its superior alignment with observed variability and temporal consistency, with correlation coefficients (R) over 0.97 and standard deviations closely matching. These results show the Residual Hybrid model's improved ability to learn both the main diurnal cycles and transient atmospheric variations, providing a strong, physically consistent tool for wind forecasting in tropical and coastal regions.