Electric vehicle charge scheduling with deep reinforcement learning agents

Abstract

With the growing market shares of electric vehicles in the automobile industry and the Icelandic government’s plans of energy transition by 2030, increasing stress will be imposed on the country’s energy grids. Iceland’s utility companies can deploy smart methods of energy allocation to electric vehicles, in stead of investing in more energy production, such as scheduling electric vehicle charging during off-peakhours.The varied behavior of electric vehicle users, their inclination to charge during peak hours and the complexity of electric vehicle charging makes the allocation of electric vehicle charge a difficult problem to solve.This work proposes the use of deep reinforcement learning as a method of optimizing the charge allocation to electric vehicle users by simulating an environment of electric vehicle users and their charging stations. Two learning agents were devised along with a greedy baseline agent for measurements. Firstly, an off-policy deep q-learning agent that learns from bootstrapping samples from the simulation, secondly, an advantage actor critic that produces a probability distribution for actionsin each state, and lastly a greedy baseline agent for measurements that charges the electric vehicles whenever possible.The results from the simulation show the possibility of deploying deep reinforcement learning methods in real world charging allocation.Keywords: Charging, Deep learning, Electric vehicle, Reinforcement learning,Scheduling, Simulation