A Multi-Agent Reinforcement Learning Approach to Price and Comfort Optimization in HVAC-Systems

This paper addresses the challenge of minimizing training time for control Heating, Ventilation, and Air-conditioning (HVAC) systems with online Reinforcement Learning (RL). is done by developing a novel approach to Multi-Agent (MARL) HVAC systems. In this paper, environment formed system formulated as Markov Game (MG) in general sum setting. The MARL algorithm designed decentralized structure, where only relevant states are shared between agents, actions sequence, which sensible from system’s point view. simulation domestic house located Denmark resemble an average house. heat source air-to-water pump, Underfloor Heating (UFH). subjected weather changes data set collected Copenhagen 2006, spanning entire year except June, July, August, not required. It shown that: (1) When comparing Single Agent (SARL) MARL, can be reduced 70% four temperature-zone UFH system, (2) agent learn generalize over seasons, (3) cost heating 19% or equivalent 750 kWh electric energy per Danish compared traditional method, (4) oscillations room temperature 40% when RL methods method.