Surrogate Estimators for Complex Bi-level Energy Management

We deal here with vehicles in charge of performing internal logistics tasks inside some protected area. Those vehicles are provided in energy by a local solar hydrogen production facility, with limited storage and time-dependent production capacities. In order to avoid importing energy, one needs to make energy production and consumption collaborate in order to minimize both production and routing costs. Because of the complexity of resulting bi-level model and because both levels may correspond to independent players, we propose a collaborative approach. Adopting the point of view of the vehicle manager leads us to short-cut the production scheduling level with the help of surrogate estimators. According to this prospect, we propose four approaches, respectively based on parametric approximation, reinforcement learning, learning through neural networks an path search in a large partially ordered set. We implement and compare them through several numerical experiments.