Robust real-time energy management for a hydrogen refueling station using generative adversarial imitation learning

As the demand for hydrogen fuel increases with the rise of fuel-cell electric vehicles (FCEVs), the energy management of hydrogen refueling stations (HRSs) is crucial for operational efficiency and environmental sustainability. Although previous studies have applied various energy management methods to HRSs, the application of data-driven approaches for real-time optimization remains very limited. This study addresses this gap by proposing a novel energy management model for optimal real-time energy scheduling of on-grid HRSs using generative adversarial imitation learning (GAIL). The proposed algorithm aims to mimic expert demonstrations to enhance decision-making. Initially, expert trajectories are constructed by collecting state-action pairs, achieved by solving a deterministic energy scheduling model using historical data and a mixed integer linear programming (MILP) solver. These expert trajectories are then used to train the GAIL algorithm. Through adversarial training involving policy and discriminator networks, GAIL accurately simulates expert behavior, enabling strategic decisions regarding power-to-hydrogen conversion, hydrogen-to-power conversion, and FCEV refueling to maximize system profit. The applicability and feasibility of the GAIL algorithm are evaluated across a wide range of scenarios. The results show that total profit increases by up to 29% with the application of the proposed GAIL algorithm. Compared to well-regarded deep reinforcement learning methods, GAIL demonstrates superior performance, proving its effectiveness in real-time energy scheduling of on-grid HRSs.