A Game‐Theoretic Reinforcement Learning Approach for Multiobjective Optimization of Building HVAC Systems Subject to Unmeasurable Disturbances

ABSTRACT Buildings remain to be the key consumers of electricity worldwide. A major contributor to this energy consumption is the building's heating, ventilation, and air conditioning (HVAC) system. Owing to the challenges of modeling large‐scale building HVAC systems, traditional model‐based optimal control approaches become increasingly difficult to use toward optimizing these systems. We present a game‐theoretic optimal control design for building HVAC systems using a two‐player non‐zero‐sum cooperative game. A data‐driven and completely model‐free Q‐learning algorithm is proposed that solves a quadratic game optimization problem online without invoking any knowledge of the building dynamics. Mass flow rate and supply air temperature are considered to be the key decision variables that serve as the two players in the game. The building HVAC zone serves as a game environment for these players, whose dynamics are assumed to be completely unknown to the players. The algorithm is shown to learn the optimal game solution and associated control policies for each player. Simulation results show the effectiveness of the proposed scheme in achieving temperature tracking under unmeasurable disturbances and unknown zone dynamics.