A surrogate machine learning modeling approach for enhancing the efficiency of urban flood modeling at metropolitan scales

Abstract

Urban drainage systems in metropolitan areas are highly complex, posing significant challenges for effective stormwater management. Traditional models like Storm Water Management Model (SWMM) are widely used but become inefficient at large scales with intricate drainage networks. This limitation is particularly critical for early warning systems, which require fast and simplified flood modeling methods. This study investigates surrogate machine learning (ML) models for efficient urban flood modeling at a metropolitan scale. Using SWMM as benchmark, the proposed model demonstrates its ability to replicate SWMM results, offering a more efficient alternative. We partition the system into hydrologically connected clusters, reducing 66,482 manholes to 363 manageable units. The approach combines this clustering strategy with ML modeling to predict key surcharge variables (flood duration, peak, and volume) for individual manholes across complex drainage system. Model validation demonstrates robust performance (R² > 0.8 for extreme events) while reducing computational time by 92.6%. Feature importance analysis reveals depth and duration as primary drivers of flood prediction, with model accuracy correlating to infrastructure density. The surrogate models excel particularly at predicting extreme events, with varying performance across different rainfall conditions. This computational efficiency enables real-time prediction updates crucial for emergency response planning and flood management strategies.