Hybrid deep learning framework for wind speed downscaling using LSTM-CNN-attention model

Accurate wind speed downscaling is crucial for optimizing wind energy production and improving turbine efficiency. This study proposes a hybrid deep learning framework combining long short-term memory (LSTM), convolutional neural networks (CNNs) and attention mechanisms to enhance local scale wind speed prediction. The proposed model is designed to capture the temporal and geographical patterns of wind speed observations and highlight the important patterns with attention mechanisms. Employing real world wind speed observations in Delhi, the paper compares the performance of the hybrid approach with conventional statistical downscaling (SD) models, machine learning (ML) models and single deep learning methods. According to the results, the LSTM-CNN-attention model has greater accuracy than other approaches, reducing root mean square error (RMSE) by ∼30% with an <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" overflow="scroll"> <mml:mtext>RMSE</mml:mtext> </mml:math> of 0.603 m/second, coefficient of determination (R2) value of 0.721, and a Pearson correlation coefficient of 0.850, indicating an exceptional improvement in downscaling accuracy. Use of spatial and temporal features enables better modeling of wind speed dynamics and the attention mechanism allows the model to concentrate more on key patterns. To our knowledge, this is the first study applying a combined LSTM-CNN-attention framework specifically for local wind speed downscaling. The findings have direct practical applications in wind turbine installation, maximum energy yield and long-term energy planning.