A comparative analysis of reinforcement learning and model predictive control for HVAC system optimization

Heating, ventilation, and air-conditioning (HVAC) systems account for an important energy de
mand and significant greenhouse gas emissions in the building sector. Traditional control 
methods often result in inefficiency and compromise occupant comfort due to their lack of dy
namic adaptability. This study compared Reinforcement Learning (RL) algorithm using the Deep 
Q-Learning (DQL) technique with Model Predictive Control (MPC) employing a Monte Carlo Tree 
Search (MCTS) algorithm. This research aimed to minimize energy consumption while providing 
thermal comfort in a building. The findings, obtained using EnergyPlus simulation indicated that 
DQL significantly outperformed MCTS-based MPC. For one year testing, DQL achieved 3.29 % 
and 4.62 % reductions in energy use compared to rule-based and on-off controls, respectively. In 
contrast, MCTS-based MPC achieved only 0.43 % and 1.8 % reductions, with higher levels of 
discomfort. DQL’s adaptability stemmed from experience-based learning and allowed effectiv