Optimizing Indoor Pollutant Exposure, Energy Consumption, and Thermal Comfort Using Deep Reinforcement Learning Agent

NISHCHAYA KUMAR MISHRA, Nipun Batra, Sameer Patel, Indian Institute of Technology Gandhinagar, India

     Abstract Number: 145
     Working Group: Reducing Aerosol Exposure with Control Technologies and Interventions

Abstract
With individuals spending nearly 90% of their time indoors, maintaining well-controlled indoor environments is critical for minimizing pollutant exposure and ensuring thermal comfort. However, achieving optimal indoor air quality often involves trade-offs between energy consumption and thermal comfort, presenting a complex multi-objective optimization challenge. Multiple data-driven machine learning and fully physics-driven algorithms have been proposed to predict indoor air quality and operate heating, ventilation, and air conditioning systems. However, their real-world integration is limited in dynamic control processes such as indoor environment control owing to modelling, computational, and learning challenges.

Reinforcement learning (RL), with its capacity for autonomous adaptation through continuous interaction with the environment, presents a promising alternative. Unlike static models, RL agents can dynamically learn control strategies with fewer prior assumptions, offering broader deployment potential. Previous studies have applied RL to optimize pollutants, energy, and thermal comfort. However, the control actions in most studies are limited to ventilation.

This work developed an RL agent to control ventilation and set temperature of the air conditioning unit to optimize exposure, energy, and thermal comfort in a house. The study evaluates the reliability of a reinforcement learning (RL) agent for indoor environmental control under varying pollutant emission scenarios by evaluating its performance against a physics-based dynamic optimization strategy. Simultaneously, the transferability of the trained RL agent is investigated across multiple residential buildings subjected to diverse ambient conditions. The results highlight the potential of RL agents as scalable and adaptable solutions for maintaining healthy indoor environments while reducing environmental impacts.