Rapid urban inundation prediction method based on numerical simulation and AI algorithm

• Published in: Journal of hydrology (Amsterdam) 2025-02, Vol.647, p.132334, Article 132334
• Main Authors: Pan, Xinxin, Hou, Jingming, Chen, Guangzhao, Li, Donglai, Zhou, Nie, Imran, Muhammad, Li, Xinyi, Qiao, Juan, Gao, Xujun
• Format: Article
• Language: English
• Subjects: algorithms; decision making; Detention Basin; drainage; Hydrodynamic Model; hydrodynamics; Link Capacity; Machine Learning; mathematical models; prediction; rain; surface water; Urban Inundation
• ISSN: 0022-1694
• DOI: 10.1016/j.jhydrol.2024.132334

•Introduces a new deep learning method for urban inundation prediction, addressing data scarcity.•Forecasts inundation events in just 27.44 seconds on average, reducing computation time.•Simulates water depth and drainage network load, offering guidance for emergency flood management. Urban inundation caused by extreme torrential rains has become one of the most prominent natural disasters globally, and rapid and precise forecasting of such events is now a primary measure in flood emergency management. However, AI-based rapid inundation forecasting requires sufficient historical inundation data, and existing forecasts only predict urban inundation without addressing elements such as the load on urban drainage systems. Therefore, this paper combines physical process models and AI technology to develop a rapid forecasting model for urban inundation, designed to quickly predict surface water accumulation, link capacity, and water depth at control nodes in storage pools due to extreme rainfall. To address the issue of insufficient historical rainfall and inundation monitoring data, the model integrates one-dimensional link network models and two-dimensional hydrodynamic models to address the shortage of flood data. The model simulates flood data for various rainfall intensities and patterns in the study area, forming a rainfall-inundation outcome matrix. This matrix is then trained using a BP neural network algorithm, ultimately producing a rapid forecasting model for urban inundation applicable to the study area. The results show: (1) In terms of computational accuracy, the predicted values for surface water accumulation, link capacity, and water depth at storage pool control nodes have R2 values of no less than 0.826, 0.951, and 0.765, respectively, demonstrating the model’s reliable prediction accuracy; (2) In terms of computational efficiency, the rapid forecasting model averages 27.44 s to forecast a single flood event, achieving a speed increase of approximately 322 times compared to traditional two-dimensional hydrodynamic models, indicating a fast computation speed. Thus, this forecasting model can provide more time for urban emergency decision-making, thereby reducing the economic losses and casualties caused by urban inundation.