Urban inundation rapid prediction method based on multi-machine learning algorithm and rain pattern analysis

Abstract

Urban inundation disasters caused by extreme rainfall events are becoming increasingly severe. However, the numerical inundation model based on physical process has relatively slow computational speed and struggle to meet the demands of current forecasting and early warning systems. Addressing how to integrate machine learning technology into rapid urban inundation prediction, especially when dealing with limited training samples, and selecting the most appropriate machine learning algorithms for different rain patterns, is a crucial and pressing issue. To tackle this problem, this work combines a coupled hydrological-hydrodynamic model known for its high-computational accuracy with efficient machine learning algorithms. Taking different rain patterns into account, a rapid urban inundation prediction method is proposed. This method utilizes data-driven results from the coupled hydrological-hydrodynamic model and constructs multiple machine learning algorithms based on the conclusions drawn from the analysis regarding the most suitable machine learning algorithms for different rain patterns. The results show that for single-peak, double-peak, and uniform rain patterns, the Ridge algorithm, KNN algorithm, and RF algorithm should be respectively employed. The method achieves a mean absolute percentage error of 5.32 %, 7.73 %, and 2.49 % for single-peak, double-peak, and uniform rainfall scenarios and can predict inundation within a 3.68 km2 area in 14.07 s. This method offers high prediction accuracy and fast computational speed, not only meeting the requirements of daily forecasting and early warning but also providing a new direction for the application of machine learning technology in urban inundation prevention.