Abstract: Hong Kong is located in the Northwest Pacific, a region frequently affected by typhoons, where extreme wind events pose significant challenges to offshore wind power development and grid stability. However, due to the relative scarcity of extreme wind speed events in observational data, the sample distribution is highly imbalanced, and traditional data-driven methods often struggle to accurately predict such events. To address this issue, this study proposes an ultra-short-term extreme wind speed forecasting method based on the synthetic minority over-sampling technique (SMOTE) combined with long short-term memory (LSTM) networks. First, SMOTE is used to generate a certain number of extreme wind speed samples, thereby alleviating the data imbalance problem and enhancing the model's sensitivity to rare events. Then, historical wind speed data are used to construct an LSTM model to capture temporal dependencies and nonlinear features. Experimental results demonstrate that, compared with the conventional LSTM model, the proposed SMOTE-LSTM approach significantly improves both regression accuracy and classification performance in ultra-short-term extreme wind speed prediction. When the number of synthetic extreme wind samples reaches 10% of the total dataset, the model achieves optimal performance: the root mean square error (RMSE) decreases from 1.58m·s^-1 to 1.55m·s^-1, the mean absolute error (MAE) drops from 1.05m·s^-1to 0.99m·s^-1, and the F1-score rises from 0.61 to 0.78. In the extreme wind speed range of 20~24 m·s^-1, the MAE is reduced from 2.29m·s^-1 to 1.52m·s^-1. This method provides a feasible approach for ensuring the reliable operation of offshore wind power systems under extreme weather conditions in the South China Sea and other typhoon-prone regions.

Key words: extreme wind, wind speed prediction, data augmentation, long short-term memory neural network