First Author: LIU Yuanyuan
Corresponding Author: LIU Yesen
Journal: Atmosphere

Abstract

Floods in hilly areas are characterized by rapid onset and destructive force. This paper introduces a novel machine learning–based method to address this problem, aiming to improve the accuracy and predictability of flash flood forecasting in small catchments within mountainous and hilly regions. The method is mainly applicable to small watersheds under 600 km2.

The proposed approach analyzes the spatiotemporal characteristics of rainfall dynamics and identifies rainfall-flood events in the history that present high similarity to current model patterns, thereby enabling real-time forecasting by “learning from the past to predict the present.” The method demonstrates significant accuracy: the average error in peak flow prediction is 8.33%, the mean error in total runoff prediction is 14.27%, and the error in peak arrival time prediction is within just 1 hour, meeting the precision requirements of flood forecasting.

Keywords: Artificial intelligence; Machine learning; Rainfall spatiotemporal characteristics; Flood risk management; Flood forecasting; LSTM neural networks