Applying Observed Rainfall and Deep Neural Network for Urban Flood Analysis

Abstract

Urban flooding primarily occurs due to flash floods in low-lying regions or insufficient capacity of drainage systems. Urban inundation has become more dangerous to residents because of an increase in impermeable areas, uncertainties in urban sewage planning, and the accumulation of sedimentation in drainage pipes. As an urban watershed with a drainage system and an urbanized river is complicated, a specific time interval is required for the preparation and processing of urban runoff analysis. To allow for rapid simulations of urban runoff estimation during heavy rainfall, a deep neural network model that imitates the conditions of a 6-h duration rainfall was developed in this study. Ten different statistical aspects for each rainfall event were considered as input data, and the total accumulated overflow from a manhole was calculated at intervals of 10 min using storm water management model (SWMM). To verify the accuracy of the results from SWMM, the results obtained from a two-dimensional hydraulic model and an inundation trace map were compared. The computational times for the deep neural network and rainfall data-based models proposed in this study were estimated to be within 1 s, whereas the simulation using SWMM required 14 min. The proposed deep learning model was tested using the total accumulated overflow for the rainfall event observed at Gangnam (400) automatic weather station on July 27, 2011. The simulated results agreed with the observed results in terms of the total accumulated discharge.