Extracting interpretable building control rules from multi-objective model predictive control data sets

Developing intelligent building control strategies is increasingly becoming a multi-objective problem as owners, occupants, and operators seek to balance performance across energy, operating expense, environmental concerns, indoor environmental quality, and electric grid incentives. Implementing multi-objective optimal controls in buildings is challenging and often not tractable due to the complexity of the problem and the computational burden that frequently accompanies such optimization problems. In this work, we extract near-optimal rule sets from a database of non-dominated solutions, created by applying multi-objective model predictive control to detailed EnergyPlus models. We first apply multi-criteria decision analysis to rank the non-dominated solutions and select a subset of consistent and plausible operating strategies that can satisfy operator or occupant preferences. Next, unsupervised clustering is applied to highlight recurring control patterns. In the final step, we build a supervised classification model to identify the right optimal temperature control patterns for a particular day. The performance of the simplified rule sets is then quantified through simulation. Despite the dramatically simpler form, the best rule sets were able to achieve 95–97% of the energy savings and 89–92% of the cost objective savings of the fully detailed model predictive controller, while achieving similar thermal comfort and peak electrical demand.