Data-Driven Predictive Control (DDPC) with Deep Neural Networks for Building Energy Savings

Building model predictive control (MPC) relies on a white- or grey-box model that can require significant time and domain knowledge to develop and calibrate, often on an ad hoc basis. Black-box models can be developed and trained with limited domain knowledge and are easily transferable. In this study, a deep neural network is trained to predict indoor temperature response using a few easily obtained predictors. The trained network is embedded within an MPC framework, replacing a grey-box model, and this data-driven predictive controller (DDPC) is implemented in a facility consisting of two side-by-side identical office spaces. One office is controlled by DDPC; the other remains under automated control. Experimental results show that DDPC reduces energy consumption by up to 30% compared to baseline control while maintaining indoor temperature throughout the day. DDPC presents a scalable solution to the challenges associated with developing and implementing building MPC on a large scale.