Deep learning-based model predictive control with exponential weighting strategy and its application in energy management systems

Building energy management plays an important role in improving the overall system efficiency and reducing energy consumption. To achieve this goal, it is significant and challenging for the optimization of energy consumption and the utilization of renewable energy sources. This work presents a deep learning-based model predictive control with exponential weighting (DLEMPC) strategy to control and optimize Energy Management Systems (EMS). First, an exponential weighting technique with decreasing characteristic is introduced to the cost function over the timeslots in the receding horizon of the MPC to improve the control performance of the system, which aims to obtain the control actions by paying more importance on recent timeslots in the finite time-horizon. Second, a controller based on the deep belief network (DBN) model is proposed to reduce computational complexity of the rolling horizon optimization in practical applications. The deep learning controller is obtained by training it with a large number of input and output data pairs that are generated from a well-defined MPC designed with the new cost function. Finally, the DLEMPC strategy is used to control and optimize an EMS, connected to a grid, battery, HVAC, and solar panel. The results demonstrate that DLEMPC strategy can significantly improve the energy efficiency of buildings and reduce energy consumption compared to the traditional MPC formula.