REinforcement learning for Microgrid Optimization and TEmperature Control (REMOTEC)

 This report documents the experimental validation of a Meta-Reinforcement Learning (Meta
RL) framework applied to building temperature control under uncertain thermal dynamics. The
 work was conducted at the SYSLAB Risø Campus, Danish Technical University, within the
 scope of the ERIGrid 2.0 Lab Access programme. The goal was to investigate how a controller
 trained in simulation, using an approximate building model, could adapt to real-world conditions
 through data-driven adaptation.
 The experimental activities encompassed initial model identification, parameter uncertainty
 analysis, simulation-based controller training, real-world data collection, and deployment of
 a learning-based adaptation mechanism. The targeted building, Power Flexhouse 03, was
 equipped with temperature sensors and actuators connected to a central control platform. Key
 findings demonstrate that the proposed Meta-RL framework substantially reduces energy con
sumption without compromising thermal comfort, with the adaptation mechanism effectively
 correcting initial modeling inaccuracies after only three days of data.
 This project shows that Meta-RL is a viable, scalable solution for adaptive control in buildings
 and can serve as a foundation for next-generation intelligent HVAC systems