Recurrent neural network training for energy management of a mild Hybrid Electric Vehicle with an ultra-capacitor

The paper exemplifies the design flow of a neural network controller for energy management of a parallel hybrid electric vehicle (HEV) with an ultra-capacitor. As the energy storage capacity of the ultra-capacitor is limited, the energy management in such a powertrain can be particularly challenging, and charging/discharging of the ultra-capacitor has to be performed optimally to reduce fuel consumption and avoid drivability degradation. In this paper, a neural network model of the powertrain is first trained from the input/output data generated by a powertrain model; with this neural network model a neural network controller, which prescribes the power split between the engine and the electric motor, is trained using a multi-stream Extended Kalman Filter (EKF)-based training method. The weights of the neural network controller are then further refined on the original plant model by a stochastic perturbation approach. The choices of the neural network architecture and of the cost function to model the plant and the controller are rationalized. The fuel consumption improvement is compared against the optimal solution obtained using dynamic programming.