Deep learning-based prediction of cold surge frequency over South Korea

Abstract

Operational seasonal prediction models have limited skill in predicting number of wintertime cold surge days over South Korea. Here, we present a hybrid prediction framework that combines a coupled general circulation model with a Long Short-Term Memory neural network. Using data from 1980/81 to 2021/22, the framework incorporates model predictions together with 24 climate indices whose robustness was confirmed through leave-one-year-out cross-validation (LOYOCV). Model performance was evaluated using both LOYOCV and an independent train-test split, ensuring robust assessment. Compared with dynamical-only predictions, the hybrid system achieved substantially higher correlation and lower root mean square error, demonstrating improved prediction skill. Shapley Additive exPlanations analysis identified key contributors such as the Scandinavia pattern, Western Pacific pattern, and Southern Oscillation Index, and regression confirmed that enhanced skill after 2000 was linked to stronger contributions from atmospheric indices. These results indicate a temporal shift in dominant teleconnection drivers from oceanic to atmospheric control, reflecting evolving sources of predictability. The framework highlights the complementary value of combining dynamical and statistical approaches. As cold surges over South Korea are embedded within larger-scale East Asian circulation anomalies, underscoring its broader regional significance and offering a pathway to improved early prediction of extremes and better climate risk management.