TR2021-132

Kernel Regression for Energy-Optimal Control of Fully Electric Vehicles


    •  Menner, M., Di Cairano, S., "Kernel Regression for Energy-Optimal Control of Fully Electric Vehicles", IEEE Vehicle Power and Propulsion Conference, DOI: 10.1109/​VPPC53923.2021.9699361, October 2021, pp. 1-6.
      BibTeX TR2021-132 PDF
      • @inproceedings{Menner2021oct,
      • author = {Menner, Marcel and Di Cairano, Stefano},
      • title = {Kernel Regression for Energy-Optimal Control of Fully Electric Vehicles},
      • booktitle = {IEEE Vehicle Power and Propulsion Conference},
      • year = 2021,
      • pages = {1--6},
      • month = oct,
      • doi = {10.1109/VPPC53923.2021.9699361},
      • url = {https://www.merl.com/publications/TR2021-132}
      • }
  • MERL Contact:
  • Research Areas:

    Control, Dynamical Systems, Machine Learning, Optimization

Abstract:

This paper presents a control algorithm for electric vehicles (EVs) with multiple motors. The control algorithm minimizes the EV's energy usage by optimizing the efficiency of its electric motors. The degrees of freedom exploited by the control algorithm are the torque-split ratio between multiple motors, the gear ratio for transmission, as well as the velocity profile of the EV. The algorithm uses kernel regression to learn a pseudo-convex cost function for optimal control from tabulated data of the electric motors' efficiency maps. The main advantages of the algorithm are its real-time feasibility due to the pseudo-convex shape and its flexible approximation capabilities. A simulation study shows how an EV with multiple but different motors and a torque-split controller can efficiently exploit the range of operation of the individual motors. The proposed algorithm achieves energy savings of up to 20% and 40% for the US06 and Urban Dynamometer Driving Schedule (UDDS), respectively, by leveraging the strengths of the different electric motors. Finally, we show that the energy-optimal velocity profile varies for different EV specifications as a result of their motor efficiencies. In particular, compared to a profile with constant acceleration, the proposed kernel regression algorithm achieves energy savings of up to 13%.