What question did this study set out to answer?

The aim is to enhance flood prediction by integrating typhoon physics into machine learning models.

April 22, 2026Open Access

Enhancing flood prediction through physics-driven typhoon feature engineering and machine learning

ZZZhi ZhangZhaoqing University YXYusha Xiao BCBiqing ChenZhaoqing University

Key Points

The aim is to enhance flood prediction by integrating typhoon physics into machine learning models.
Developed a machine learning framework for the Boluo watershed incorporating typhoon feature engineering.
Evaluated four models: Linear Regression, Artificial Neural Network, Random Forest, and XGBoost under various feature scenarios.
Applied SHAP analysis to understand feature contributions during extreme events.
Achieved superior performance with ANN-EnTY showing a Kling-Gupta Efficiency of 0.946 and RMSE of 174 m³/s.
Demonstrated a 3.1% improvement in KGE and a reduction of 16.7 m³/s in RMSE from the Baseline scenario.
Reduced peak prediction error by approximately 4% during an extreme flood event with a peak flow of 7670 m³/s.

Abstract

Typhoon-induced extreme floods pose severe threats to subtropical watersheds, yet systematic integration of typhoon physics into machine learning flood prediction remains limited. This study developed a physics-informed machine learning framework for the Boluo watershed, South China, emphasizing typhoon feature engineering. Four models (Linear Regression (LR), Artificial Neural Network (ANN), Random Forest (RF) and XGBoost (XGB)) were systematically evaluated across three feature engineering scenarios: Baseline (conventional hydrometeorological variables), With Typhoon (original typhoon observations), and Enhanced Typhoon (19 physics-informed derived features). Physics-driven design included sigmoid-transformed distance decay functions representing saturating near-field typhoon influence, multi-day cumulative impact indices integrating antecedent storm effects, and trajectory-based kinematic features characterizing translation speed and directional evolution. ANN-EnTY achieved superior performance with Kling-Gupta Efficiency (KGE) of 0.946 and Root Mean Square Error (RMSE) of 174 m³/s, representing a 3.1% improvement in KGE and 16.7 m³/s reduction in RMSE compared to the Baseline scenario. During a representative extreme flood event with peak flow of 7670 m³/s, ANN-EnTY reduced peak prediction error by approximately 4% relative to the best-performing baseline model. SHAP analysis revealed upstream flow dominance (72.6%), while typhoon features, contributing only 2% overall, played critical synergistic roles during extremes. This dual-mode pattern of routine memory-driven versus extreme event-driven responses provides mechanistic insights for operational flood warning systems. The framework offers replicable methodology for typhoon-prone watersheds with direct implications for disaster preparedness and water management.

Ask AI

Helpful

Bookmark

View Full Paper