Center for Research in Energy Systems Transformation (CREST)

Model uncertainty is a significant challenge to more widespread use of model predictive controllers (MPC)for optimizing building energy consumption. This paper presents two methodologies to handle modeluncertainty for building MPC. First, we propose a modeling framework for online estimation of states andunknown parameters leading to a parameter-adaptive building (PAB) model. Second, we propose a robustmodel predictive control (RMPC) formulation to make a building controller robust to model uncertainties.The results from these two approaches are compared with those from a nominal MPC and a commonbuilding rule based control (RBC). The results are then used to develop a methodology for selecting acontroller type (i.e. RMPC, MPC, or RBC) as a function of building model uncertainty. RMPC is found to bethe superior controller for the cases with an intermediate level of model uncertainty (30–67%), while thenominal MPC is preferred for the cases with a low level of model uncertainty (0–30%). Further, a commonRBC outperforms MPC or RMPC if the model uncertainty goes beyond a certain threshold (e.g. 67%).