Surrogate-based multiobjective optimization to rapidly size low impact development practices for outflow capture

Abstract

The optimal implementation of low impact development (LID) practices remains an open research issue. Multiobjective optimization (MOO) based on physically-based models facilitates LID sizing but requires large computational budgets. We developed and validated a surrogate MOO framework based on machine learning (ML) models to obtain satisfactory optimal solutions with an affordable computational cost. First, a calibrated and validated storm water management model (SWMM) was driven by a 2-year, 120 min Chicago pattern storm with random LID areas for generating 10,000 samples. Second, ML models, including multiple linear regression (MLR), generalized regression neural network (GRNN), and backpropagation neural network (BPNN), were trained and validated to predict catchment total outflow (CTO) and catchment peak outflow (CPO). Third, using the outperformed ML models as surrogates, the LID areas were optimized by non-dominated sorting genetic algorithm-II to minimize construction costs, CTO, and CPO. The results obtained for an urban catchment in Fengxi, China, demonstrated that: (1) BPNN outperformed MLR and GRNN in terms of accuracy, while the training time (∼40 min) was ∼6857 times and ∼706 times that of MLR (∼0.35 s) and GRNN (3.4 s) models, respectively. (2) The optimization process using BPNN models as surrogates converged well, indicating the efficiency of the proposed framework. (3) Under the marginal effect, the permeable pavement was identified as the leading practice for most Pareto solutions, followed by bioretention cell and green roof. Moreover, a subcatchment with a higher runoff coefficient was deemed preferable for LID implementation. (4) Outflow reduction by the LIDs was significant (CTO and CPO were reduced by 28.6–56.9 % and 27.5–48.3 %, respectively) but negligible for retarding the time of CPO. (5) The optimization time was reduced by ∼95.82 % using the surrogate-based MOO instead of the SWMM-based model when the training time of the BPNN models was ignored. This study provides insights into the efficient optimization of LIDs.