Operations of Fuel Cell Vehicle-to-Grid Systems: From Rule-based to Supervised Learning

Abstract

To combat rising greenhouse gas emissions in the transportation sector, hydrogen-powered fuel cell electric vehicles (FCEVs) present a promising alternative. A key advantage of this technology is the ability to use FCEVs as mobile power generation devices in vehicle-to-grid (V2G) stations. This combination of fuel cell electric vehicle-to-grid (FCEV2G) was found to be economically viable, but the profits depend on the station's operation strategy. This thesis investigates an optimal operation strategy for an FCEV2G station, which progresses from baseline operational simulation to an advanced intelligent agent model developed using machine learning. First, a detailed rule-based operational simulation was developed for a FCEV2G station using historical data from Ontario as an example. This model was improved from the literature by incorporating several key real-world components, including the hydrogen cycle, dynamic FCEV participation patterns, and variations in individual FCEV efficiency due to pre-existing degradation. The analysis concluded that the station’s operational performance is limited because it is a rule-based operation stategy that is unable to act optimally within any given hour. This limitation can be explained in three operational failures. First, its non-optimal dispatch logic causes the system to fail to reserve its limited hydrogen for periods of peak value. Second, this mismanagement of hydrogen is then amplified by low round-trip efficiency. Finally, the station’s operation is constrained from using high-cost market hydrogen to buffer this deficit. To overcome these limitations, the second phase focused on developing a machine learning (ML) agent. A behavioral cloning agent was trained to mimic the decisions of an expert i.e., a Mixed-Integer Linear Program (MILP) optimizer, which establishes the theoretical profitability of the system. The trained agent demonstrated definitive success, achieving 93.2% of the expert's optimal profit on training data and a robust 80.4% on an unseen test set. This high performance confirms the feasibility of using ML agent for the FCEV2G operation. This approach also provides a significant advantage in decision speed: the agent makes decisions in milliseconds, replacing the computationally intensive MILP expert. Analysis of the agent's behavior revealed that it successfully learned to navigate volatile market conditions, including extreme price shocks, by mastering the expert's forward looking strategies. In conclusion, this research delivers a successful proof-of-concept for an intelligent FCEV2G operational controller. The primary contribution is demonstrating that a fast ML agent can learn the forward-looking operational strategies of a slow optimizer. By mastering these strategies, the agent achieves near-optimal profitability in real time, proving a viable pathway for deploying intelligent control systems to manage the day-to-day operations of volatile energy assets.