Power management in hybrid electric vehicles using deep recurrent reinforcement learning

A power management framework for hybrid electric vehicles (HEVs) is proposed based on deep reinforcement learning (DRL) with a Long Short-Term Memory (LSTM) network to minimize the fuel consumption through determining the power distribution between the two propulsion sources, the internal combustion engine (ICE) and the electric motor (EM). DRL is effective for handling the high-dimensional state and action spaces in the HEV power management problem, and the LSTM structure leverages temporal dependencies of input information, providing internal state predictions automatically without introducing extra state variables. This technique is entirely online, meaning that the framework is constructed in real time during the training phase, independently of a prior knowledge of driving cycles. The learned information stored in the LSTM network is utilized efficiently, and the computational speed is enhanced by making multiple predictions simultaneously in each step. Simulation over various driving cycles demonstrates the efficacy of the proposed framework in fuel economy improvement.