Abstract:

Bayesian optimization (BO) has recently been demonstrated as a powerful tool for efficient derivative-free optimization of expensive black-box functions, such as those prevalent in performance optimization of complex energy sys- tems. Classical BO algorithms ignore the relationship between consecutive optimizer candidates, resulting in jumps in the admissible search space which can lead to fail-safe mechanisms being triggered, or undesired transient dynamics that violate operational constraints. In this paper, we propose LSR-BO, a novel global optimization methodology that enforces local search region (LSR) constraints by design, which restricts how much the optimizer candidate can be changed at every iteration. We demonstrate that naively incorporating LSR constraints into BO causes the algorithm to get stuck in local sub- optimal solutions, and overcome this challenge through the development a novel exploration strategy that can gracefully navigate the trade-off between short-term “local”, and long- term “global”, performance improvement. Furthermore, we provide theoretical guarantees on the convergence of LSR-BO. Finally, we verify the effectiveness of our proposed LSR-BO method on an illustrative benchmark and a real-world energy minimization problem for a commercial vapor compression system.