Parallel Monte Carlo-based surrogate optimization of building energy models

Abstract

Surrogate optimization aims to optimize black-box functions that are expensive to evaluate, usually in terms of long evaluation times. The goal is to reach a promising region of the design space in as few evaluations as possible, so that no timelimit is imposed on the optimization step used to select the next evaluation point. EnergyPlus in recent versions became a multi-threaded application, making it possible to simulate several instances of a parametric building model in parallel on different processor cores. Here we propose a conceptually simple methodology to perform parallel surrogate optimization of building energy models, whose simulation results are treated as black-box functions. It relies on gradient-boosted regression tree ensembles (XGBoost) to model black-box functions that may depend on either real, integer or categorical parameters. It uses Monte Carlo-based optimization of surrogate model to quickly approximate Pareto front between the predicted function values and an uncertainty metric. Candidate designs are divided into groups with low, medium and high uncertainty, and the next simulation sample is then selected as a combination of best predicted function values in the low uncertainty group, uniform sample of Pareto solutions in the medium uncertainty group, and the most uncertain candidates in the last group. We illustrate the proposed method on surrogate optimization of both synthetic test functions and the energy performance of a residential villa in Kuwait. Favorable behavior of the proposed method is observed in comparison with well-known parallel surrogate optimization methods, as well as genetic algorithms.