Data-driven Koopman model predictive control for the integrated thermal management of electric vehicles

Abstract

The thermal management system (TMS) of electric vehicles (EVs) consumes a considerable amount of energy, and hence its optimal control is crucial for enhancing EV driving range. However, the complexity of an integrated TMS and its varying operation modes bring challenges for real-time optimal control. The assumptions and simplifications adopted for developing computationally inexpensive physics-based control-oriented models often result in prediction errors. To address the impact of model errors, this study proposes a Koopman-based model predictive control (MPC) approach for the integrated TMS operation in EVs, which includes a cooling mode change. Koopman prediction models are developed based on the Extended Dynamic Mode Decomposition (EDMD) structure utilizing data collected from high-fidelity MATLAB/Simulink® simulations. For the selection of Koopman models, a corrected Akaike Information Criterion ($AI{C}_c$) is applied to thirteen candidates. In addition, the prediction performance of the selected models is evaluated by examining open-loop simulation errors during the cooling mode change with different prediction lengths. These selected Koopman models are then implemented in a Quadratic Programming (QP)-based MPC structure. The corresponding controllers are integrated into the high-fidelity MATLAB/Simulink® plant model and evaluated under four driving conditions. Compared with a nonlinear MPC (NMPC) baseline controller addressing the same optimal control problem, the chosen Koopman controller demonstrates improved temperature regulation performance and a 6.5% reduction in energy consumption. The Koopman controller reduces the computational time for each calculation, decreasing from 247 ms to 54 ms, compared to the NMPC controller.