Energy Optimization in Smart Buildings Using Deep Q-Network-based Reinforcement Learning

Abstract

The rapid expansion of smart building technologies demands innovative solutions to optimize energy consumption while preserving comfort. Traditional rule-based and supervised learning approaches often lack adaptability to dynamic environmental conditions, leading to inefficiencies in HVAC and lighting control. Reinforcement learning (RL) offers a promising alternative by enabling autonomous, data-driven decision-making in complex building environments. This study proposes a Deep Q-Network (DQN)-based RL framework for real-time energy management in smart buildings. The system integrates real-time sensor data (temperature, occupancy, weather) with a virtual building model (EnergyPlus + OpenAI Gym) to train an adaptive control agent. A custom reward function balances energy savings and thermal comfort, while experience replay stabilizes training. The framework was evaluated against rule-based and supervised learning baselines using metrics such as energy consumption (kWh), comfort deviation (ASHRAE standards), and control stability. The proposed system achieved a 22% reduction in energy consumption compared to conventional rule-based systems while maintaining a significantly lower comfort violation rate of just 5%, outperforming traditional methods that exhibited a 12% violation rate. The reinforcement learning approach demonstrated superior adaptability to dynamic occupancy changes and weather fluctuations, though this enhanced performance came with inherent trade-offs between computational cost and real-time responsiveness that must be carefully considered in practical implementations. These results demonstrate the system’s ability to optimize both energy use and comfort under real-world conditions. The results also validate RL as a scalable solution for sustainable building operations, bridging the gap between simulation and real-world deployment.