Machine Learning Improves Assessment of Mesoscale Offshore Wind Resources over the Yellow Sea

Numerical weather prediction (NWP) models, widely applied in mesoscale offshore wind resource assessment, face inherent limitations in accuracy. This study explores the potential of machine learning (ML) techniques to enhance the NWP model performance for offshore wind resource assessment. Three ML models are proposed to correct biases in wind field simulations generated by the Weather Research and Forecasting (WRF) model, utilizing in-situ observations as references. The findings reveal systematic biases in seasonal wind field simulations over the Yellow Sea in China, with overestimation in winter and underestimation in summer. Initial WRF simulations result in a mean root mean square error (RMSE) of 3.15 m s^-1 and a Pearson correlation coefficient (PCC) of 0.76. After ML correction, all three models show significant forecasting improvements. Specifically, the hybrid model incorporating convolutional neural networks (CNN) with long short-term memory networks (LSTM) and attention mechanisms (CNN-LSTM-Attention) achieves the greatest enhancement, reducing RMSE to 1.48 m s^-1 and increasing PCC to 0.95, representing improvements of 53.1% and 25.3% (p < 0.05), respectively, compared to the original WRF model. Similarly, the extreme gradient boosting (XGBoost) and deep neural network (DNN) models demonstrate substantial improvements, achieving RMSE reductions of 45.4% and 51.3%, and PCC increases of 22.7% and 24.6% (p < 0.05), respectively. Notably, XGBoost performs well under stable wind conditions, whereas the CNN-LSTM-Attention model excels at capturing dynamic wind speed variations and processing long-sequence datasets, showcasing superior generalization on independent test sets. This study highlights the significant potential of ML in improving NWP-based wind field simulations, providing a robust scientific foundation for precise offshore wind resource assessments.