Comparative evaluation of deep learning and statistical models for wind speed downscaling

Accurate wind speed downscaling is essential for meteorological applications and reliable station-scale wind estimation. This study presents a systematic comparison of recurrent deep learning architectures, including RNN, GRU, LSTM, Bidirectional LSTM, Sequence-to-Sequence LSTM, and Stacked LSTM, against conventional statistical models for wind speed downscaling using paired IMD station observations and NASA MERRA-2 reanalysis data for Pune, India. All models were evaluated under identical preprocessing and training protocols to ensure a fair comparison. Among the standard recurrent architectures, the Stacked LSTM achieved the best predictive performance. An Attention-Enhanced LSTM was subsequently evaluated as an extension to investigate the benefits of incorporating a soft-attention mechanism. It achieved the highest overall accuracy, with an RMSE of 0.5034 and an MAE of 0.3650, demonstrating additional performance gains over the standard recurrent architectures. The findings provide a robust benchmark for deep learning-based wind speed downscaling and future atmospheric prediction studies.