Machine Learning Based Charging Decision Policy for a Fleet of Electric Vehicles
Examensarbete för masterexamen
In a fleet of autonomous electric vehicles, each vehicle must frequently charge its battery. The charging of the vehicles should ideally be as efficient as possible without causing roadblocks or battery depletion. The aim of this thesis, which is performed in collaboration with Volvo Autonomous Solutions, is to develop such charging policies using reinforcement learning methods. More precisely, the problem is formulated as a Markov decision process (MDP), and tabular Q-learning and deep Q-learning are used to learn (optimal) charging policies. The learned policies are evaluated and compared against each other, as well as against a rule-based policy used as a benchmark. The robustness and generalizability of the learned policies are tested by adjusting a number of site-specific parameters, such as charging time and number of vehicles at the site. The results indicate that machine learning-based charging policies can increase the efficiency for a fleet of vehicles compared to a simple rulebased charging policy. Moreover, the importance of such policies is also shown to vary depending on the site configuration and its complexity.
charging policy, machine learning, reinforcement learning, Q-learning, DQN, deep Q-learning, Markov decision process