Online learning for distributed optimal control of an electric vehicle fleet

The management of electrical power systems requires the resolution of large-scale problems whose agents are linked by coupling constraints. Nevertheless, decomposition methods cannot provide an exact solution while dealing with temporal dynamics in a stochastic environment. Each agent would then have to solve a local problem in which future quantities intervene. However they depends on other agents' future decisions which are still unknown. In order to enhance the existing approximate approaches to this problem, the proposed method involves Alternating Direction Method of Multipliers to overcome the large dimension by an iterative resolution of local coordinated problems. Uncertain temporal dynamics are handled by a stochastic dynamic programming approach. In order to make local problems tractable, an online learning step is added. The agents can then anticipate future global variations in a local but uncertain way. The optimal charging of an electric vehicle fleet paired with a wind power plant is considered as a case study. The expected benefits are highlighted, both at the outset and after training the anticipatory models. The discussion addresses the learning parameters allowing the fastest convergence.