Comparative study of various typical hydrological models for flash flood forecasting

ABSTRACT Flash floods in mountainous regions pose significant threats, yet comparative studies of different modeling approaches under consistent conditions remain limited. This study systematically compares four hydrological modeling approaches representing distinct philosophies for flash flood forecasting across 12 mountainous watersheds in China: parameter estimation-based flash flood model, storage-infiltration compatible model, Hydrologic Engineering Center's Hydrologic Modeling System, and long short-term memory neural network. Using 236 historical flood events with a 70%/30% calibration–validation framework, we evaluate model performance through prediction accuracy, temporal precision, and computational characteristics. Results reveal distinct performance patterns: Hydrologic Engineering Center's Hydrologic Modeling System achieved highest Nash–Sutcliffe efficiency (0.783), followed by storage-infiltration compatible (0.782), long short-term memory (0.745), and parameter estimation-based flash flood model (0.740). For peak flow prediction, parameter estimation-based flash flood model demonstrated highest qualification rates (87.03%), followed by long short-term memory (84.04%), storage-infiltration compatible (79.6%), and Hydrologic Engineering Center's Hydrologic Modeling System (79.12%). Peak timing accuracy showed comparable performance across models (0.89–1.04 h: average error).The comparative analysis reveals model-specific strengths: physically-based models excel in overall hydrograph simulation, conceptual models provide balanced performance, data-driven approaches show efficient pattern recognition, while parameter estimation methods demonstrate advantages in peak flow prediction. Study provides objective benchmarks for flash flood forecasting.