Abstract:

Physics-informed simulation models of heating, ventilation, and cooling (HVAC) systems play a critical role in predicting system dynamics and enabling analysis, control, and optimization of buildings and equipment. The predictive performance of these simulation models are strongly linked to calibration mechanisms: algorithms that systematically select parameter values that optimize a given calibration-cost map (e.g., L-2 error). Poorly selected parameter values typically result in large deviations between measured building data and simulated data, limiting the utility of the simulation model in subsequent design. State-of-the-art calibration methods explore the parameter space by computing numerical gradients that are susceptible to measurement noise or employing population-based search mechanisms that require exorbitant data. To improve robustness and curtail data requirements, one can ‘learn’ or approximate the calibration-cost map and subsequently leverage the topology of the approximated function to find good search directions despite noisy measurements. Concretely, we employ machine learning to construct a calibration-cost map to direct model calibration for systems with joint dynamics of buildings and HVAC equipment. The learner explores subregions of the parameter space with high uncertainty and queries the model only where collecting simulation data yields useful information. This leads to lower simulation data-requirements compared to widely used calibration mechanisms.