This study presents a hybrid rainfall prediction framework, TCR-BNN, which integrates a reduced-physics Tropical Cyclone Rainfall (TCR) model with a Bayesian Neural Network (BNN) to predict tropical cyclone (TC) rainfall with both improved accuracy and uncertainty quantification. Trained on 150 TC events from the North Atlantic basin (2000–2024) and evaluated on an additional 40 storms, TCR-BNN consistently outperforms the physics-based TCR model and a pure data-driven BNN. TCR-BNN benefits from TCR’s physics-based rainfall intensity estimates and the BNN’s ability to learn complex spatial rainfall patterns from atmospheric variables, while also providing probabilistic uncertainty estimates. The hybrid approach achieves an R-squared of 0.86 for accumulated rainfall within a 200 km radius of the storm center and accurately captures heavy rainfall from intense hurricanes. The TCR-BNN model offers a practical and computationally efficient approach toward more reliable TC rainfall prediction.
Weng et al. (Wed,) studied this question.