Data-Driven Design of Model Predictive Control for Powertrain-Aware Eco-Driving Considering Nonlinearities Using Koopman Analysis

Eco-driving is a highly nonlinear control problem. The nonlinearities include the complex energy conversion/dissipation in the powertrain, environmental influences such as road grade and aerodynamic drag, constraints due to traffic signs, safety issues, and physical limits of the vehicle system. In recent years, researchers have increasingly revisited the Koopman operator to linearize nonlinear dynamics. This paper adopts such an approximation technique to construct the lifted state space in a data-driven procedure that allows us to incorporate nonlinearities and system perturbations in the cost function. In addition, the nonlinear constraints in states can also be handled linearly. The resultant formulation of a linearly constrained quadratic program can be readily applied to design a model predictive control that enjoys a low computation load as with a linear dynamic system. Simulation results demonstrate additional energy saving potential compared to a linear approach.