Data-driven surrogate modeling: Introducing spatial lag to consider spatial autocorrelation of flooding within urban drainage systems

Abstract

Data-driven surrogate modeling has been increasingly employed for flooding simulation of urban drainage systems (UDSs) due to its high computational efficiency and accuracy. However, spatial autocorrelation is prevalent in many typical scenarios, including the UDS. This omission of spatial information is very likely to cause the machine learning model to capture the wrong UDS overflow mechanism from the data. To capture the spatial autocorrelation, an artificial neural network (ANN)-based surrogate modeling method that introduces spatial lag to account for the spatial autocorrelation of flooding within the UDS is proposed and coupled with a genetic algorithm (GA) to reduce the uncertainty caused by random initialization of ANN. In this study, a surrogate modeling experiment was carried out for the Storm Water Management Model (SWMM). The experimental results show that the ANN can successfully capture the spatial autocorrelation induced by flooding within the UDS and accurately replicate the output simulated by SWMM.