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. Indeed, each agent would have to solve a local minimisation in which future quantities intervene. However, these quantities depend on other agents' future decisions which are still unknown. In order to enhance the existing approximate approaches to this challenge, the proposed method involves Alternating Direction Method of Multipliers to overcome the large dimension by an iterative resolution of local coordinated minimisations. Uncertain temporal dynamics are handled by a stochastic dynamic programming approach. In order to make local problems solvable, the online learning of a Markov process is added. The agents can then anticipate future global variations in a local probabilistic 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 anticipative models. The discussion addresses the learning parameters allowing the fastest convergence.