Learning Near-optimal Decision Rules for Energy Efficient Building Control

Recent studies suggest that advanced optimization based control methods such as model predictive control (MPC) can increase energy efficiency of buildings. However, adoption of these methods by industry is still slow, as building operators are used to working with simple controllers based on intuitive decision rules that can be tuned easily on-site. In this paper, we suggest a synthesis procedure for rule based controllers that extracts prevalent information from simulation data with MPC controllers to construct a set of human readable rules while preserving much of the control performance. The method is based on the ADABOOST algorithm from the field of machine learning. We focus on learning binary decisions, considering also the ranking and selection of measurements on which the decision rules are based. We show that this feature selection is useful for both complexity reduction and decreasing investment costs by pruning unnecessary sensors. The proposed method is evaluated in simulation for six different case studies and is shown to maintain the high performance of MPC despite the tremendous reduction in complexity.