Wind power sequences under extreme weather conditions exhibit strong non-stationarity and abrupt variation characteristics. Meanwhile

the scarcity of historical samples makes it difficult for traditional deep learning methods to be fully trained

leading to a significant decline in prediction accuracy. To address this issue

this paper proposes a two-phase adaptive forecasting framework based on a pre-trained time-series foundation model

termed the Typhoon two-phase adaptive prediction model (TTPAPM). First

a typhoon adaptation module is designed to separately represent the periodic components and abrupt disturbance components of the input sequence

thereby enhancing the model’s perception capability for abnormal meteorological features. Then

the two types of features are concatenated and fused before being fed into the pre-trained time-series foundation model to achieve hourly-scale power prediction. Finally

a two-stage hybrid optimization strategy is introduced to effectively correct the data distribution shift caused by extreme weather through targeted calibration of both regular and typhoon distributions. The results show that: 1) The proposed method can be used to effectively solve the problems of the modeling difficulty caused by sample scarcity under typhoon conditions

and compared with the best baseline model

the coefficient of determination (R²) under regular weather and typhoon weather is improved by 10.5% and 3.5%

respectively; 2) Under the zero-shot setting

the prediction accuracy under regular and typhoon weather reaches 0.85 and 0.71

respectively

indicating good transfer ability and prediction stability; 3) In terms of computational resources

the memory consumption is significantly lower than that of the baseline models. This study can provide an effective reference for applying time-series foundation models to wind power prediction under extreme weather scenarios.