Surrogate Global Optimization for Identifying Cost-Effective Green Infrastructure for Urban Flood Control With a Computationally Expensive Inundation Model

Abstract

Optimization algorithms and urban inundation models are powerful tools to identify cost-effective designs of urban green infrastructures such as low-impact developments (LIDs). Most previous LID design optimization studies are based on one-dimensional (1D) inundation models, which cannot provide spatial information of flooding. The LID design optimization on two-dimensional (2D) models or coupled 1D-2D models was rarely explored due to the expensive computing time. This work investigates the effectiveness of surrogate optimization methods for LID design, which have not been used for the LID design problems. We propose a general LID design optimization framework that searches the optimal LID configurations based on both the spatial flood damage under a series of probable flood events and the life cycle costs (LCCs) of LID. We demonstrate the framework using a case study for an urban catchment with 55 sub-catchments and 103 LID decision variables. We tested two different surrogate optimization methods designed for high-dimensional problems: DYnamic COordinate search using Response Surface models (DYCORS) and Trust Region Bayesian Optimization (TuRBO), and one popular non-surrogate method particle swarm (PSO). The result indicates that: (a) DYCORS is a promising method (significantly faster than TuRBO and PSO) for identifying the optimal LID design to minimize the flood damage cost and LID LCC; (b) Optimized LID design could reduce damage cost by as much as $12.14 million for the urban catchment after eliminating its own LCC compared with no LID implementation; (c) LID is effective in reducing the imperviousness of lands in urban areas.