Using Machine Learning to Identify and Optimize Sensitive Parameters in Urban Flood Model Considering Subsurface Characteristics

Abstract

This study presents a novel method for optimizing parameters in urban flood models, aiming to address the tedious and complex issues associated with parameter optimization. First, a coupled one-dimensional pipe network runoff model and a two-dimensional surface runoff model were integrated to construct an interpretable urban flood model. Next, a principle for dividing urban hydrological response units was introduced, incorporating surface attribute features. The K-means algorithm was used to explore the clustering patterns of the uncertain parameters in the model, and an artificial neural network (ANN) was employed to identify the sensitive parameters. Finally, a genetic algorithm (GA) was used to calibrate the parameter thresholds of the sub-catchment units in different urban land-use zones within the flood model. The results demonstrate that the parameter optimization method based on K-means-ANN-GA achieved an average Nash-Sutcliffe efficiency coefficient (NSE) of 0.81. Compared to the ANN-GA and K-means-deep neural networks (DNN) methods, the proposed method better characterizes the runoff generation and flow processes. This study demonstrates the significant potential of combining machine learning techniques with physical knowledge in parameter optimization research for flood models.