Real-Time Optimization of Control Strategy for a Range-Extended Electric Vehicle using Reinforcement Learning Algorithm and Neural Network 2020-01-1190
Range-Extended Electric Vehicles (REEV) have seen an increase in market share in the past decade. This trend can be attributed to an increased market shift towards electrified powertrains while addressing the range anxiety usually associated with an electric vehicle. In such a scenario, operating the vehicle efficiently is critical to meet the CAFÉ standards.
This energy optimization problem becomes even more critical if the vehicle is being operated as part of a fleet as minimal energy savings get compounded across the fleet and result in significant savings for the service provider and more affordability for the customers. There is also an upward trend in ride sharing services operated by fleet owners like Uber and Waymo. Fleet vehicles offer the unique advantage of availability of large amounts of data about the consumer usage pattern in a given area. When coupled with traffic density and immediate destination of the current consumer of the vehicle, the data can assist the improvement of fuel economy while a traditional rule-based strategy can hardly take advantage of the data.
Deep Orange 11 program at Clemson University is focusing on solving the mobility needs in the year 2035. In the powertrain of the REEV, a battery acts as the primary energy source and an internal combustion engine (ICE) acts as a range extender. Optimizing the energy usage of such a vehicle is a critical aspect of the problem statement of this program.
This paper utilizes an approximated Reinforcement Learning based approach as the energy management strategy for such a fleet-based vehicle. More specifically, the controller follows the model-free Q-learning approach to interact with vehicle and optimize the energy management strategy in real time. The state-action function is approximated by Neural Networks.
A high-fidelity vehicle simulator is developed in Simulink to validate the strategy and demonstrate the potential savings against a general rule-based strategy which is currently most commonly used in the industry. At the end of learning, the proposed approximated Reinforcement Learning achieves significant fuel saving compared to the rule-based strategy over the driving cycle.
Nipun Mittal, Aditya Pundlikrao Bhagat, Shubham Bhide, Bharadwaj Acharya, Bin Xu, Chris Paredis