The charging queue can change the energy system’s forecast

A shared electric-and-hydrogen refuelling hub has a coordination problem hiding in its queue: the order in which vehicles receive service helps determine how much electricity and hydrogen the station will need. In year-long case studies, a new artificial-intelligence system that learns both the queue and the energy response achieved the lowest reported vehicle-service delay and energy cost among 11 methods (Yao et al., 2026). The paper’s supplied text does not report the size of that advantage, but its central idea is important: the energy system should not treat demand as a forecast arriving from somewhere else. The station’s own scheduling decisions create that demand.

That is a practical distinction. A hub serving battery-electric vehicles and hydrogen vehicles must manage chargers and refuellers alongside solar panels, batteries, an electrolyser that makes hydrogen, hydrogen storage, and connections to the electricity grid and hydrogen pipeline. A decision about who gets served now changes both the waiting line and the load the energy equipment must meet. The researchers’ answer is to give the two jobs to separate learning agents—and to train them in a way that accounts for how one agent changes the world experienced by the other.

The Science

The paper studies an electric-hydrogen-transport system, or EHTS: a transport-service operation joined to a local energy system. Electric vehicles draw power from chargers. Hydrogen vehicles receive fuel through refuellers. On the supply side, solar generation can serve charging demand directly, charge a battery, or run an electrolyser to produce hydrogen. Stored energy can later help meet demand; the grid and a hydrogen pipeline cover what local resources cannot.

The operation proceeds in two steps during each decision interval. First, vehicle scheduling (VS) determines which waiting vehicles have priority and which can be assigned to available service equipment. Then energy dispatch (ED) decides how to supply the resulting electricity and hydrogen demand. The order matters: the scheduler has already shaped the demand by the time the dispatcher acts. The energy dispatcher does not, in turn, change which vehicles were assigned during that same interval.

This is different from a familiar energy-management setup in which demand is supplied as an input—perhaps a forecast or a fixed profile. Here, the demand is endogenous: it emerges from decisions made inside the system. If an assignment allows chargers to serve vehicles for more of an interval, electricity demand rises accordingly. The corresponding service decisions also determine how much hydrogen the refuellers dispense. That demand is then passed to the energy agent.

There is a second complication: the two decisions have different shapes. Vehicle scheduling is a combinatorial problem. With several candidate vehicles, the number of possible priority orders grows rapidly, and a policy that directly chooses a vehicle-to-charger assignment would also need to cope with the number and arrangement of chargers. Energy dispatch is instead a continuous-control problem: the agent must choose amounts of power and hydrogen to move among equipment and storage. One action is an ordering; the other is a set of energy-flow settings.

The authors therefore build a heterogeneous pair of reinforcement-learning agents. Reinforcement learning is a way of improving decisions through repeated interaction with an environment and feedback about outcomes. For vehicle scheduling, the system uses proximal policy optimisation (PPO), a method for updating a decision policy, together with a Plackett–Luce policy. The latter represents a probability distribution over rankings: it lets the agent learn which vehicle should be placed earlier in a priority order without enumerating every possible permutation.

The scheduler does not directly assign vehicles to individual chargers or refuellers. Instead, it ranks candidate electric and hydrogen vehicles separately. A deterministic greedy allocator then translates those rankings into feasible assignments, choosing compatible equipment with the earliest feasible service start. Vehicles already connected stay in service until their current demand is met; they are not interrupted mid-session. If a service finishes before the interval ends, the equipment can serve another vehicle. This design keeps the learned decision focused on priority while leaving the physical allocation rule explicit.

The energy agent uses twin delayed deep deterministic policy gradient (TD3), a reinforcement-learning method designed for continuous actions. It controls battery and hydrogen-tank operation and how surplus solar power is allocated. Solar power is first used to serve electric-vehicle charging demand; any surplus can charge the battery, supply the electrolyser, or be curtailed. The model does not allow grid electricity to power the electrolyser: hydrogen production is treated specifically as a way to use surplus renewable generation.

The agents’ actions have a directional relationship. Scheduling determines demand for the dispatcher, but dispatch does not determine the already-made schedule for that same interval. Yet the learning process creates a feedback problem over time. As the scheduler improves, it can produce a different pattern of electricity and hydrogen demand. The dispatcher’s stored experience—past examples of states, actions, and outcomes—may then represent an older version of the scheduler rather than the current one.

To address that mismatch, the researchers propose three-stage training. The agents first undergo layer-wise pretraining, allowing each to learn its part of the problem before joint operation. They then fine-tune concurrently. During that phase, the ED agent’s replay capacity and update intensity are reduced, so it relies less heavily on old experiences as scheduling behaviour changes. The goal is not to make the agents’ decisions identical or fully independent, but to let the dispatcher adapt to the demand patterns the scheduler is currently creating.

What They Found

The abstract reports the headline result: across year-long case studies, the proposed framework achieved the lowest vehicle-service delay and energy cost among 11 evaluated methods (Yao et al., 2026). That is a notable pairing. A scheduler that reduces waiting might increase the energy cost if it concentrates demand at expensive times or consumes stored energy that would be more valuable later. Conversely, a strategy that economises on energy could leave vehicles waiting. The reported result suggests that coordinating the two decisions can improve both measures relative to the alternatives tested.

But the available text does not include the results section, numerical tables, or the values plotted in the paper’s charts. It therefore does not support a defensible graph of the performance differences, nor does it tell us how large the improvement was, which methods made up the full set of 11, or whether the best result held under every tested condition. No chart is included here rather than inventing measurements or turning a ranking into a fabricated margin.

The paper says its evaluations vary solar photovoltaic penetration, vehicle arrival intensity, and charger and refueller capacity. Those are meaningful stressors. More solar changes how much renewable electricity is available; busier arrivals can make queues and service bottlenecks more severe; and equipment capacity constrains how many vehicles can be served. The authors present these tests as a way to examine generalisation beyond the training configuration. Since the supplied text does not provide the detailed results, however, it is not possible to say how performance changed across those scenarios or how robust the gains were.

The study also gives a useful methodological result even apart from the headline ranking: it makes the link between scheduling and dispatch explicit in the learning architecture. The first agent’s output becomes the second agent’s demand input, rather than the dispatcher being trained against an assumed independent demand profile. That is the key modelling move behind the reported coordination.

Why This Changes Things

The problem resembles a small but increasingly important version of a larger infrastructure challenge. Transport is becoming more electrified, while some heavy-duty or specialised uses may rely on hydrogen. Their energy needs meet at depots, stations, ports, logistics centres, and other sites where the vehicles’ operating schedules and the energy system’s constraints are inseparable. A hub that treats vehicle activity as fixed demand may miss an operational lever: the order and timing of service can shape when local energy resources are used.

This does not mean that a learning algorithm can simply make a busy station efficient. The operational details matter. The paper assumes grid and pipeline imports have enough capacity to cover residual demand. That assumption makes the supply problem tractable, but real sites may face connection limits, pipeline constraints, or high costs during particular periods. The model also simplifies transitions between vehicles at the same pile by neglecting the time needed to disconnect one vehicle and prepare for the next. Such details can influence real-world throughput.

The split between learned priorities and deterministic allocation is also a pragmatic design choice. Directly learning every individual assignment would create a much larger action space, and could tie the policy to one particular number of chargers or refuellers. A priority ranking is easier to represent and can, in principle, travel across facility layouts. But the greedy allocator still governs how those priorities translate into service. The learned ranking is not a complete station-control policy by itself; performance depends on the interaction between the policy and that allocation rule.

The training issue is especially relevant beyond this case study. In many connected systems, one algorithm changes the data another algorithm sees. A delivery scheduler changes the workload faced by a warehouse; a building controller changes the demand seen by a grid operator; a charging policy reshapes the load that an energy manager tries to smooth. Experience that was useful yesterday can become less representative after another policy changes. The paper’s staged training approach is one way to recognise that learning agents do not operate in a static environment simply because the physical equipment is unchanged.

There are also reasons to read the reported result as a promising simulation finding, not a deployment guarantee. The abstract provides no numerical effect sizes or uncertainty estimates in the text supplied here. We cannot assess whether the cost and delay improvements are large enough to justify implementation, how they compare with a carefully optimised non-learning baseline, or how the learned policies behave when arrivals and energy prices depart from the test conditions. Nor does the material provided establish performance in a live station with human operators, equipment faults, or imperfect forecasts.

What’s Next

The next test is whether the coordination survives contact with real operating constraints. A field evaluation would need to measure not only average delay and energy cost, but also queue extremes, missed service targets, equipment utilisation, renewable-energy use, and the frequency with which grid or pipeline imports approach their limits. It should report comparisons in absolute units as well as percentages, so an operator can judge whether a gain is meaningful at a particular site.

A stronger evidence base would also show the full set of benchmark results and scenario-by-scenario performance. The paper’s abstract says the study examines differing solar penetration, arrival intensity, and service capacity. Publishing those values would clarify whether the dual-agent system is consistently better or whether its advantage depends on particular station conditions. It would also help identify when a simpler scheduling rule or a more conventional optimiser might be sufficient.

Finally, the learning architecture raises a design question for future work: how should a station balance the flexibility of learned decisions with guarantees about service and safety? Priority policies could be assessed against explicit rules for maximum waiting times, critical fleet commitments, or limits on energy imports. And if other parts of the system become decision-makers—such as price-responsive charging, hydrogen production from the grid, or fleet operators with their own priorities—the tidy one-way coupling could become a more complex negotiation.

For now, the paper’s clearest contribution is conceptual and practical: it treats the queue and the energy system as parts of one operation, while respecting that their decisions are different and occur in sequence. The reported best-in-class results among 11 methods make that approach worth testing further. If the gains hold under transparent numerical comparisons and real infrastructure limits, coordinated learning could help multi-energy hubs serve vehicles more promptly while using their local energy resources more economically.