A Fast Adaptive Temperature Control Approach for Uncertain Building Models via Meta-RL

This work presents a novel approach to temperature control in buildings using Meta-Reinforcement Learning (Meta-RL) to address uncertainties in building thermal dynamics. The proposed method utilizes a control-oriented model to train a Meta-RL controller in simulation across a wide range of building dynamics, generated by varying the most influential parameters of the building dynamics within large uncertainty ranges. As a result, the Meta-RL controller learns adaptable control strategies tailored to the dynamic characteristics within the whole uncertainty region. Initially, an RL agent is employed to create a suitable encoding of thermal dynamics parameters. During the deployment, an Adaptation Module substitutes the RL agent inferring the right encoding from real-time data. The paper provides details on the specifically designed modeling and control architecture and its parametrization. Extensive validation demonstrates that the developed Meta-RL architecture is able to effectively and quickly adapt to diverse building dynamics, is as efficient as a standard dedicated MPC control, and finally is also able to quickly restrict the uncertainty ranges during deployment.