An Adaptive Model-Predictive Control Informed Rule-based Control for Residential Cooling Operations under Extreme Weather Events

Due to air conditioner (AC)/heat pump unit fault, cooling capacity degradation usually exists in residential buildings. This sometimes makes the heating, ventilation, and air-conditioning (HVAC) system struggle in maintaining thermal comfort when the cooling load is getting higher, especially in extreme weather events (e.g., heat waves). A potential solution to tackle these issues is to pre-cool the house to deal with the limitation of house thermal mass or the HVAC capacity. Adaptive model-predictive control (AMPC) is a promising control strategy that can help mitigate those over-heating risks. It predicts the future system state and makes appropriate actions through optimizations ahead of time, and thus effectively responds to disturbance variations (e.g., weather). Optimization is essential to AMPC, but matrix multiplication and inversion cannot be executed online in most existing building controllers. Therefore, AMPC informed rule extraction is proposed to extract simple operation rules from the results of a large-scale AMPC offline implementation. Then, those computationally efficient simple rules will be online executed in the residential HVAC controller. The goal of rule extraction is to select the minimum sets of inputs and feed them to the rule-based controls (RBC) and maintain the same or close levels of energy consumption and thermal comfort with the AMP simultaneously. The AMPC-informed RBC avoids the online execution of computationally expensive optimization, thus alleviating computation load in building HVAC controllers. It is anticipated that if rules are extracted with large sets of data covering different operation scenarios and climate zone, these heuristic rules can be used by practitioners and homeowners, even without a learning model-based adaptive control framework. This paper implements an AMPC in a virtual testbed of a DOE prototype EnergyPlus residential building model. Classification and Regression Tree (CART) are adopted as the rule extraction algorithm to extract operation rules from the AMPC results. The test case is for RBC with an AC unit cooling capacity degradation under an extreme weather event of heatwaves. A preliminary comparison of the performance of ad-hoc RBC, AMPC, and AMPC-informed RBC strategies will be presented.