Robust Typhoon Rainfall forecasting based on machine learning and Bayesian model averaging

Abstract

Typhoon-induced heavy rainfall can lead to severe flooding, causing significant damage to social systems. Although machine learning (ML) offers an efficient approach for typhoon rainfall forecasting, individual models often exhibit considerable uncertainty. To address this, this study proposes a robust typhoon rainfall forecasting model based on Bayesian Model Averaging (BMA) with four ML models including Random Forest (RF), Support Vector Regression (SVR), K-Nearest Neighbors Regression (KNN), and eXtreme Gradient Boosting (XGBoost) across 0–6 h lead time. The model incorporates three types of features within the typhoon impact area (radius of 400 km): typhoon characteristics, grid spatial attributes and meteorological characteristics. Typhoon characteristics of each grid are dynamically weighted to reflect the weakening typhoon impacts with increasing distance from the typhoon center. Based on these features, multiple scenarios consisting of two experiments (Input Unchanged (IU) and Rolling Forecast (RO)) paired with various input designs (ALL: all features, SHAP3: 3 most critical variables, and LAG4: 4 lagged rainfall) are designed to explore the performance of all models. Case study of 28 typhoons affecting Guangdong Province from 2020 to 2023 clearly demonstrated that the weighted typhoon wind speed contributed the most in typhoon rainfall forecasting, and the BMA approach significantly enhanced the forecast accuracy. Leveraging sufficient effective information as model inputs could significantly improve the predictive performance of individual models in long-term forecasting. This method could provide flexible and suitable scenario options for data-rich and data-scarce regions, supporting early disaster warning during typhoon-prone seasons.