Physics-informed linear regression is competitive with two Machine Learning methods in residential building MPC

Because physics-based building models are difficult to obtain as each is individual, there an increasing interest in generating suitable for MPC directly from measurement data. Machine learning methods have been widely applied this problem and validated mostly simulation; are, however, few studies on a direct comparison of different or validation real buildings be found the literature. Methods that indeed application often lead computationally complex non-convex optimization problems. Here we compare physics-informed Autoregressive-Moving-Average with Exogenous Inputs (ARMAX) Learning based Random Forests Input Convex Neural Networks resulting convex schemes experiments practical goal minimizing energy consumption while maintaining occupant comfort, numerical case study. We demonstrate Predictive Control general leads savings between 26% 49% heating cooling energy, compared building's baseline hysteresis controller. Moreover, show all model types satisfactory control performance terms constraint satisfaction reduction. However, also see ARMAX lower computational burden, superior sample efficiency models. even if abundant training data available, significantly prediction error than models, which indicates encoded prior former cannot independently by latter.