Stochastic Control of Electric Vehicle Charging

Kyle Anderson CS 229 Machine Learning Final Project Abstract— This project attempts several methods to optimize charging schedules for electric vehicles on a constrained radial network using machine learning. In the first approach, electric vehicles act as independent agents in a QLearning framework, receiving negative rewards based on congestion charges calculated from their contribution to overloaded parts of the network. In this model, electric vehicles may take the action of increase or decrease the charging rate by a fixed amount at each time interval, but are subject to charging limits, battery capacity limits, and charging deadlines. Next, I consider a central aggregator who implements a supervised learning technique to recognize the discounted future cost of a network state. During the testing phase, the controller will greedily take actions that move to a better state. Finally, I consider another central aggregator scheme based on Linear Quadratic Regulation. I propose an (experimental) alteration to the LQR method to accommodate exponential action spaces by adding a polynomial number of quasi-time steps to allow the decision maker to perform many actions within a given time step.