Experimental Evaluation of Data-Driven Predictive Indoor Thermal Management

This paper considers the problem of thermal management in a typical shared indoor space that may be equipped with multiple heterogeneous heat sources and have different temperature requirements in different sections (thermal zones) of the shared space. Utilizing an on-campus smart conference room as a testbed, we discuss the practical challenges involved in real-time data-driven model learning, when a simple first-order dynamical model is used to capture the dependencies between the heat controls and the air temperatures measured at sensor locations. The data-driven model is then utilized for predictive control of the thermal environment towards minimizing the error between the desired and attained temperatures, and the integrated solution is evaluated against a standard thermal control employed by the BMS.