Overcoming the Energy–Health Trade-Off in Smart Buildings: AI-Driven Dynamic Thresholding for Sustainable HVAC (Heating, Ventilation, and Air Conditioning) Actuation and Multi-Pollutant Management
TL;DR
An intelligent, energy-efficient, and robust multi-pollutant forecasting and control framework that integrates hybrid LSTM–GRU forecasting with Proximal Policy Optimization (PPO)-based reinforcement learning to maintain indoor air quality while minimizing unnecessary energy consumption and mechanical actuation is developed.
Abstract
Modern smart buildings face the challenge of balancing energy-saving requirements with strict indoor air quality regulations. The aim of this research is to develop an intelligent, energy-efficient, and robust multi-pollutant forecasting and control framework that integrates hybrid LSTM–GRU forecasting with Proximal Policy Optimization (PPO)-based reinforcement learning to maintain indoor air quality while minimizing unnecessary energy consumption and mechanical actuation. Using Green Computing principles, the agent significantly reduces excessive mechanical energy use while dynamically and simultaneously optimizing five pollutant dimensions (NH3, NO2, CO, PM2.5, and O3). The agent’s decision-making is guided by a new sustainable multi-objective reward function that balances energy efficiency (γ), mechanical stability (β), and health and safety (α). To evaluate the robustness and practicality of the suggested RL agent, a series of highly stressful simulated scenarios was used along with an empirical physical environment. In these stress tests, the system was subjected to dynamic environmental anomalies, including sudden weather-related temperature spikes and high-occupancy conditions, resulting in localized spikes in pollution (CO and PM2.5) and measurement corruption due to hardware malfunctions and sensor noise. Experiments have demonstrated that the dynamic framework maintains a much higher mean control threshold (0.88) than conventional static baselines (0.47), resulting in an estimated 77.3% reduction in HVAC-related energy demand based on the analytical HVAC energy model introduced compared with the strict static baseline. Moreover, the agent simultaneously outperformed the static controller’s compliance range of 12% to 17.6%, achieving a robust 80% multi-pollutant compliance rate, even under extreme simulated anomalies, including severe occupancy-driven emissions spikes and sensor network failures.