Abstract:

A critical component of model predictive control (MPC) in building energy management systems is the ability to reject exogenous disturbances such as occupant-induced heat loads, appliance loads, ambient temperature, and solar radiation. During the design phase of model predictive control, engineers typically select a limited set of disturbance scenarios through handcrafting or sampling from simple distributions. However, this approach often fails to capture a representative set of scenarios that elicit diverse closed-loop behaviors across the full operational envelope. This work addresses this limitation by proposing a combinatorial multi-armed bandit (CMAB) framework for systematically discovering representative disturbance scenarios using real building data. We formulate the scenario selection problem as a diversity maximization task, where the reward function quantifies the behavioral diversity of closed-loop responses through information-theoretic criteria such as dynamic time warping distance. The proposed approach treats the building simulation environment as a black-box system, making it applicable to complex, proprietary, or non-differentiable building models commonly encountered in practice. To address data scarcity challenges, we extend the framework by incorporating time-series generative models, specifically diffusion-based networks, to synthetically augment limited real datasets. Experimental validation using a commercial net-zero energy building demonstrates that synthetic data augmentation significantly enriches the diversity of discovered scenarios compared to using real data alone, as evidenced through principal component analysis and uniform manifold approximation projections. The CMAB algorithm successfully identified representative scenarios that revealed controller vulnerabilities not detected by conventional selection methods, leading to practical improvements in HVAC system design. The approach scales linearly with the number of scenarios and bandit iterations, making it computationally feasible for grid-interactive building energy management applications.