Energetic Optimal Control Of A Dwelling With Multiple Thermal Zones Through Deep Reinforcement Learning

Abstract

The implementation of deep reinforcement learning (DRL) has significantly advanced across various scientific fields, overcoming many inherent difficulties. However, specific challenges have emerged in each area. In the control of HVAC systems in buildings, scalability limitations have been identified, hindering its application in environments with multiple thermal zones or numerous agents. To address this issue, this work presents a control method for multiple agents in multiple thermal zones of a dwelling. This method facilitates scalability by implementing a control policy based on a deep neural network with fully shared parameters, used by all agents. This application represents the state of the art in fully cooperative multi-agent systems, ensuring effective communication among agents for optimal control of the dwelling. The implementation of this method in a social housing unit in the province of Mendoza demonstrates its effectiveness in complex scenarios. The limitations encountered are discussed, and future research directions are suggested.