Energy-Efficient Indoor Air Quality Control Via Cloud-Based Reinforcement Learning and Modular IoT Sensors

Indoor air quality (IAQ) plays an important role in human health, comfort, and cognitive performance. This work presents an edge-cloud control framework that integrates reinforcement learning (RL) with low-cost air quality sensors for real-time ventilation control. A mass-balance simulation of multiple pollutant measures (AQI, TVOC, and eCO2) was used to train two RL agents, Deep Q-Network (DQN) and Proximal Policy Optimization (PPO) via a two-phase curriculum that first learns pollutant regulation and later incorporates energy efficiency. In simulation, both agents achieved stable AQI control. DQN provided the best energy-performance trade-off$(142\ \text{Wh})$, reducing energy use by approximately 39 % compared to the rule-based baseline (234 Wh) while maintaining acceptable pollutant levels. PPO delivered the strongest pollutant tracking but at the highest energy cost$(375\ \text{Wh})$. The rule-based controller achieved low overall pollutant values but at a higher cost compared to DQN. All controllers were then deployed in the cloud and connected to a Raspberry Pi5 via MQTT. The Pi collected real-time sensor data and controlled a 775DC motor fan during alcohol-induced pollutant spikes. Although pollutant dynamics were influenced by uncontrollable outdoor conditions, both RL agents showed robust responsiveness, increasing air abstraction during spikes and reducing it afterward to save energy.