Development of a Physics-Guided Neural Network Model for Effective Urban Flood Management

Abstract

Urban flooding is a common disaster occurring every year, leading to the loss of lives and properties throughout the world. Its frequency and severity have increased over the years and are expected to increase further due to climate change impacts. Therefore, a robust urban flood management framework is needed to mitigate and/or lessen the adverse impacts of urban floods, which requires flood models capable of quickly producing accurate flood forecasts at key locations in an urban area. Previous attempts at urban flood modeling have focused on the use of physics-based rainfall-runoff models or data-driven models in isolation, with reasonable accuracy. In this study, a novel physics-guided neural network modeling approach is proposed that is capable of exploiting the advantages of both the physics-based and data-driven techniques. We employed the MIKE FLOOD model as the physics-based rainfall-runoff model and artificial neural networks (ANNs) as the data-driven technique in this study. The climatic, hydrologic, and physiographic data derived from the Indian Institute of Technology (IIT) Kanpur, a small urban area in Northern India, were employed to develop and test the proposed models and methodologies. The results obtained in this study demonstrated that both the MIKE FLOOD and ANN models employed in this study were able to represent the rainfall-runoff behavior of IIT Kanpur catchment very well. The performance of the ANN models was found to be better than that of the MIKE FLOOD models based on the performance evaluation measures considered in this study. A close examination of the results around the peak of the flood hydrographs revealed that all the models presented here performed much better around the flood peak than in the rest of the flood hydrograph, which is an encouraging result. It was found that the ANNs are powerful tools capable of producing accurate flood forecasts quickly under extreme weather conditions, and therefore they can be employed as surrogate models to achieve greater efficiency in urban flood management activities.