Improving Tropical Cyclone-Induced Rainfall Forecasting for Vietnam Using Pattern-Matching Algorithms and Selective Ensemble

Abstract. Tropical cyclones (TCs) can have devastating effects on lives and communities, making accurate rainfall forecasts essential for disaster preparedness. This study introduces a selective ensemble approach aimed at improving precipitation forecasts for TCs affecting Vietnam. The approach selects the most representative members from the ECMWF ensemble by comparing them with GSMaP observational data, using decision trees combined with various pattern-matching machine learning algorithms, including statistical techniques, K-means clustering, and hierarchical agglomerative clustering (HAC). A case study of Typhoon MOLAVE (October 27, 2020) over 12 and 24-hour accumulated rainfall shows that K-means and HAC methods showed significant improvements in replicating observed rainfall patterns, particularly in high-rainfall zones and areas with strong gradients. To evaluate robustness, the method was tested across all TCs in the East Vietnam Sea between 2017 and 2020, showing that selective ensemble approaches outperformed the raw ensemble in 70–90 % of cases, depending on the method and forecast lead time. The most notable improvements were observed when fewer than 15 ensemble members were selected, particularly for 12-hour rainfall accumulations. Categorical forecast evaluations further confirm that all selective methods yielded better performance than the raw ensemble in different rainfall thresholds, with higher improvement in 12-hour rainfall accumulations, while the improvement for 24-hour rainfall accumulations is limited. For probabilistic forecasts, selective methods generally outperformed the raw ensemble for 12-hour accumulated rainfall, whereas the improvement for 24-hour accumulated rainfall is more modest. These findings suggest that the proposed method can have some added value in improving short-term TC-induced rainfall forecasts.