Informing Building Retrofits Using Surrogates of Physics-Based Simulation Models: A Comparison of Multi-Objective Optimization Algorithms

Surrogate models are increasingly used to reduce the computational costs of building performance simulation (BPS) models. However, they are rarely coupled with optimization algorithms to inform retrofitting decisions. The goal of this paper is to compare the performance of several commonly used metaheuristic algorithms on a machine learning (ML) surrogate of a BPS model to better shape best practices in future work on surrogate model optimization. A surrogate model with an hourly prediction scheme is developed, and retrofits are optimized to minimize annual electric and natural gas loads, as well as PMVSET as a measure of discomfort. This study incorporates exploratory landscape analysis (ELA) to better understand the problem, then separately applies four optimization algorithms (random search, OMOPSO, NSGA-II, AMOSA). NSGA-II is the best-performing algorithm, finding 74% of the final set of Pareto efficient points and converging in 9% of the time required by the second fastest algorithm.