Reinforcement Learning based Energy Management of Multi-Mode Plug-in Hybrid Electric Vehicles for Commuter Route

Optimization-based (OB) methods used in vehicle energy management strategies (EMSs) have the potential to significantly increase fuel economy and extend the electric-only range of plug-in hybrid electric vehicles (PHEVs). However, OB methods are difficult to apply to current real-world vehicles because accurate detailed and high-resolution information about the future, including second-by-second vehicle velocity trajectory data, are not currently available in the current transportation infrastructure. In this paper, a practical reinforcement learning (RL) algorithm for automatic mode-switching of a multimode PHEV is introduced. The PHEV used in the work was a 2016 Chevrolet Volt driven on a simulated commuter route. The goal is to blend the charge depleting and charge sustaining modes during the trip to reduce gasoline consumption and extend electric-only range. The RL algorithm was first trained offline on recorded trips and then used in real-time when the vehicle was driven on a new trip of the same route. While OB methods like dynamic programming can find globally optimal solutions given complete information about a future trip, the RL method developed here does not require detailed future trip information and still obtains substantial improvements. Results show that the fuel economy on a miles per gallon equivalent (MPGe) basis was improved between 5.5% and 6.4% for a tested commuter route as a function of starting battery state of charge using the developed algorithm. The developed method provides an immediate solution to extend electric-only range in PHEVs used on daily commuter routes.