Practical deployment of reinforcement learning for building controls using an imitation learning approach

• Real implementation of imitation learning for a DRL controller managing a TABS. • DRL reduces energy use by 40 % while mimicking RBC with behavioural cloning. • DRL requires 45 % less energy than PI in a digital twin. • DRL reduces temperature violations by 43 % and 13 % compared to RBCs in a digital twin. This paper addresses the critical need for more efficient and adaptive building control systems to maximise occupant comfort while reducing energy consumption. Our objective is to explore the practical application of model-free Deep Reinforcement Learning (DRL) in real-world building environments by developing a system that learns and adapts to changing conditions, beginning its operation by imitating an existing Rule-Based Control (RBC) system. This approach ensures initial reliability and performance while setting the stage for advanced learning capabilities. The methodology involves two distinct phases. Initially, the DRL controller mimics the behaviour of the RBC system, using imitation learning with behavioural cloning as a safe and efficient strategy to achieve baseline operational efficiency. Subsequently, the controller is implemented within a real building in an online learning setting. In this phase, the controller utilises real-time data to continuously refine its control policy, responding adaptively to occupant behaviours and external environmental conditions. To validate our approach, we conducted a comprehensive analysis, comparing the performance of our DRL controller against the baseline RBC controller, another RBC, and a PI (Proportional-Integral) controller implemented in a digital twin model of the real office environment. Energy consumption and temperature violations related to a temperature acceptability range are considered as metrics, providing a robust framework for assessing the effectiveness of our system. The results indicate that our DRL controller, supported by imitation learning, outperforms the two RBCs by reducing energy consumption by 40 % while reducing the cumulative sum of temperature violations by 43 % and 13 % with respect to the two RBCs. Although the PI controller ensures better performance in terms of temperature violations compared to DRL, it requires 45 % more energy than the proposed DRL controller due to its inherent inability to deal with multi-objective control problems. In conclusion, this paper demonstrates the feasibility and advantages of implementing advanced DRL techniques in real-world building control scenarios. Integrating imitation learning with a DRL controller offers a novel and effective way to enhance the scalability of DRL systems, expanding their application in buildings and driving significant improvements in energy efficiency.