Cascaded reinforcement learning based supply temperature control

Abstract

In this work, a Q-learning based supply temperature control approach for a demonstrator building is proposed. The purpose is to improve the temperature behaviour inside the building and to tackle comfort problems such as overheating and undercooling which cannot be coped with by the standard heating curve. The Q-learning controller considers predicted future weather data as system states. This can be shown to be superior to Q-learning controllers without weather prediction. Furthermore, in order for the controller to capture different thermal effects of different time constants, a cascaded control structure is designed: An inner Q-learning based controller deals with thermal effects of smaller time constants. It is wrapped by an outer slower Q-learning controller which can tackle effects of larger time constants. Therewith, further improvement of comfort can be achieved.