The operation of fuel cell electric vehicle-to-grid (FCEV2G) stations presents a significant challenge due to the need to manage onsite hydrogen production, storage, and vehicle dispatch in volatile electricity markets. While mixed-integer linear programming (MILP) can determine the theoretically optimal, perfect-foresight strategy, its high computational cost makes it unsuitable for real-time deployment. This study proposes and validates a machine learning (ML) framework to create a rapid and effective operational controller. We first developed a sophisticated MILP expert, enhanced with a rolling horizon and a lookahead mechanism, to generate a dataset of optimal decisions using historical Alberta market data. Subsequently, an agent with deep neural network policy was trained using behavioral cloning to mimic this expert. The agent’s architecture was designed with a 48-dimensional observation space, critically incorporating multi-horizon price forecasts to enable strategic foresight. The trained behavioral cloning agent demonstrated exceptional performance, capturing 93.2% of the expert’s maximum possible profit on the in-sample training data and a robust 80.4% on an out-of-sample test set, confirming its ability to generalize. Further analysis revealed that the agent successfully learned a smart but cautious operational strategy, characterized by a conservative inventory policy and forecast-driven patience. In conclusion, this research delivers a successful proof-of-concept for an intelligent FCEV2G operational controller. The primary contribution is the validation of a method to replace the complex, long-term strategy of a slow optimizer with a fast, near-optimal ML policy capable of making decisions in milliseconds. This work demonstrates a viable pathway for deploying data-driven, real-time operation systems for complex energy assets.
Cetin et al. (Mon,) studied this question.