Performance Assessment of Data-Driven and Physical-Based Models to Predict Building Energy Demand in Model Predictive Controls

Abstract

The implementation of model predictive controls (MPCs) in buildings represents an important opportunity to reduce energy consumption and to apply demand side management strategies. In order to be effective, the MPC should be provided with an accurate model that is able to forecast the actual building energy demand. To this aim, in this paper, a data-driven model realized with an artificial neural network is compared to a physical-based resistance–capacitance (RC) network in an operative MPC. The MPC was designed to minimize the total cost for the thermal demand requirements by unlocking the energy flexibility in the building envelope, on the basis of price signals. Although both models allow energy cost savings (about 16% compared to a standard set-point control), a deterioration in the prediction performance is observed when the models actually operate in the controller (the root mean square error, RMSE, for the air zone prediction is about 1 °C). However, a difference in the on-time control actions is noted when the two models are compared. With a maximum deviation of 0.5 °C from the indoor set-point temperature, the physical-based model shows better performance in following the system dynamics, while the value rises to 1.8 °C in presence of the data-driven model for the analyzed case study. This result is mainly related to difficulties in properly training data-driven models for applications involving energy flexibility exploitation.