Energy Demand Prediction: A Partial Information Game Approach

This article proposes an original approach to predict the electric vehicles (EVs)’ energy demand in a charge station using a regret minimization learning approach. The problem is modelled as a two players game involving: on the one hand the EV drivers, whose demand is unknown and, on the other hand, the service provider who owns the charge station and wants to make the best predictions in order to minimize his regret. The information in the game is partial. Indeed, the service provider never observes the EV drivers’ energy demand. The only information he has access to is contained in a feedback function which depends on his predictions accuracy and on the EV drivers’ consumption level. The local/expanded accuracy and the ability for uncertainty handling of the regret minimization learning approach is evaluated by comparison with three well-known learning approaches: (i) Neural Network, (ii) Support Vector Machine, (iii) AutoRegressive Integrated Moving Average process, using as benchmarks two data bases: an artificial one generated using a bayesian network and real domestic household electricity consumption data in southern California. We observe that over real data, regret minimization algorithms clearly outperform the other learning approaches. The efficiency of these methods open the door to a wide class of game theory applications dealing with collaborative learning, information sharing and manipulation.