Physics-Enhanced LSTM for Multi-Scale Offshore Wind Power Forecasting in Typhoon-Prone South China Sea

To address insufficient prediction accuracy for the Huarui SL1500-82 offshore wind farm under complex conditions in the South China Sea, this study proposes a hybrid framework integrating aerodynamic parameters and a hyperparameteroptimized LSTM network [1]. Key features include the wind shear exponent ($\alpha$) and hub-height wind speed ($w s_{\text {hub }}$); Bayesian optimization is used to tune hyperparameters such as the learning rate (0.003) and time step (72 hours), enhancing the capture of nonlinear mappings. Using 2014-2022 data for training, 2023 data for validation, and 2024 data for testing, the model achieves a root mean square error (RMSE) of $\mathbf{4 0. 2 ~ k W}$ and a coefficient of determination ($R^{2}$) of $\mathbf{0. 9 4}$. T-tests ($p\lt 0.01$) demonstrate its superiority over ARIMA (RMSE $67.8 \mathrm{~kW}, R^{2} \mathbf{0. 7 7}$) and Random Forest (RMSE $52.1 \mathrm{~kW}, R^{2} 0.86$). Hourly forecasts for 2025-2026 capture seasonal ($\mathbf{2 7 \%}$ in summer, $43 \%$ in winter) and diurnal patterns (peaking at 14:00-16:00, troughing at 3:00-5:00) with a historical deviation of $\lt 4 \%$, facilitating effective dispatch [2].