Lifted Graph-based Modeling for Linear Predictive Control of Nonlinear Energy Systems

This paper presents a method for creating lifted linear approximations of nonlinear graph-based models. While graph-based modeling has been successfully applied as a modular and scalable approach for control-oriented representation of energy dynamics, the governing equations often include nonlinearities that present challenges to closed-loop control. The proposed method systematically approximates a linear graph from an existing physics-based nonlinear graph by lifting the model with additional vertices and edges learned from data. Numerical results for a single-phase thermal management system illustrate that, as compared to the conventional linearization method of a first-order Taylor series, the proposed method shows superior performance in both open-loop simulation and closed-loop model predictive control.