Researchers certify safe RL framework, cutting HVAC energy use 27.6%

HVAC systems account for the largest share of building energy use, yet traditional controllers struggle to balance comfort and efficiency across multiple zones simultaneously. Deep reinforcement learning offers adaptive, data-driven control — but neural network policies have lacked formal safety guarantees, limiting real-world deployment. A new paper from researchers including Oussama Ziadi and Adnane Saoud (arXiv:2608.17235) tackles this head-on by combining RL training with post-training safety certification.

The team trained PPO and Soft Actor-Critic (SAC) agents in an EnergyPlus/Sinergym simulation to minimize energy consumption while preserving thermal comfort in an eight-zone variable refrigerant flow (VRF) residential testbed. After training, they formalized safety guarantees on the PPO policy using Lipschitz-based forward invariance analysis, a technique that bounds how quickly a neural network's outputs can change — ensuring it will respect temperature constraints even in untested states. Over an annual simulation, PPO achieved a 67% reduction in comfort violations compared to rule-based control, while SAC delivered 27.6% energy savings. The PPO policy passed formal certification with a safety margin of 2.003°C. Accepted to the IEEE Conference on Control Technology and Applications (CCTA) 2026, the work demonstrates that rigorous safety verification and reinforcement learning are not mutually exclusive — and could unlock trustworthy AI-driven building automation.