Hardware-in-the-loop test of learning-based controllers for grid-supportive building heating operation

Abstract

While MPC is the state-of-the-art approach for building heating control with proven cost savings and improvement in energy flexibility, in practice, buildings are operated by simple rules-based controllers which are not able to accomplish an energy efficient and flexible operation. This paper explores the suitability of deep neural networks for approximating optimal economic MPC strategies for this task. In particular, we develop a convolutional neural network controller and test it in a closed-loop simulation against MPC and an improved predictive rule-based controller. The learned controller is easy to implement and fast to process on standard building control hardware. The feasibility, performance and robustness of the learned controller is validated in a realistic hardware-in-the-loop test setup for the demand-responsive operation of a heat pump combined with a storage tank.