Hybrid Electric Vehicle Energy Management Using ECMS, Dynamic Programming and Reinforcement Learning

How should a parallel hybrid car split its power between engine and motor? This project compares a rule, ECMS, adaptive ECMS and a DQN agent against the dynamic-programming bound on six official EPA drive cycles, in simulation on an own backward P2 model. On held-out cycles adaptive ECMS stays within 3.4 % of the bound and the tuned rule within 6.2 %; the DQN agent is behind both on all test cycles (seed means +7.0 to +28.2 %), and many of its seeds do not sustain charge. These are simulation results, not measured vehicle data.

Category: Automobile Engineering

Automobile Engineering Project AU05 from ProTechSC Academy, with graphical abstract, video demonstration and project guidance.