Control of HVAC-Systems Using Reinforcement Learning With Hysteresis and Tolerance Control

This paper presents the idea of using tolerance control in Deep Reinforcement Learning to improve robustness and reduce training time. This paper is a continuation of I where it is shown that Reinforcement Learning (RL) can be used to control an underfloor heating (UFH) system. However, it is seen in the study that the initial training time is too high and that the performance during training is not fulfilling the requirements to a UFH system. In this paper the fundamental challenge regarding control of UFH systems is explained, how RL can be beneficial for control of UFH systems, and how the implementation is done. Furthermore, results are presented with a standard hysteresis control, an RL control, and an RL control with tolerance control. These results show that the effect of tolerance control in these types of systems is significant. Finally, we discuss the challenges there are for a real-world implementation of RL-based control in UFH system.