Multi-zone field study of rule extraction control to simplify implementation of predictive control to reduce building energy use

Abstract Model-based predictive control (MPC) has been promoted as a software-based energy saving approach for buildings since the 2000s. However, there has not been significant industry adoption of MPC using current approaches. This paper proposes and tests an approach to MPC using rule extraction (RE) that can be easily implemented in building controllers to override sub-optimal control. The resulting decision trees could be implemented by building control programmers to save energy when ambient conditions are predicted to satisfy the thermal requirements of the spaces. A detailed MPC algorithm using inverse models was implemented in 27 rooms of an institutional building to provide data for a classification learning approach. Cooling and heating season decision trees were generated based on the inputs and outputs of the detailed MPC algorithm. An ensemble and sample randomization were used to generalize the trees across rooms and prevent overfitting to individual rooms. Cooling season MPC and RE energy savings were 42% and 27% respectively and heating season MPC and RE energy savings were 18% and 33% respectively. Both algorithms also reduced the difference between the temperature setpoint and measured indoor air temperature compared to the reactive controls.