The Charging Station Was Built for a Driver Who Does Not Exist

Every EV charging station in America was placed by a model. And every one of those models assumed you would charge your car the same way.

Not the way you actually do. The way the model needed you to behave in order to make the math work.

That gap — between how people actually charge and how the infrastructure assumes they will — is the subject of a new 56-page survey paper from researchers at William & Mary, and it may be the most consequential invisible assumption in the energy transition.

The paper, submitted to arXiv on May 20, 2026, introduces what the authors call the PSB trilemma: the observation that EV charging network design requires simultaneous decisions across three layers — Planning (where to build), Scheduling (when and how much power to dispatch), and Behavior (how drivers actually choose stations and charging times) — and that all three cannot be optimized together at full fidelity. Any integrated approach must degrade at least one layer. The math will not permit all three at once.

We have been building the infrastructure anyway.

The trilemma is not a discovery in the sense of a new empirical finding. It is a diagnostic — a synthesis of more than 200 prior studies that shows the fragmentation in the field is not accidental but structural. The authors show that most published work fixes one or two layers exogenously and optimizes the third, because optimizing all three simultaneously is computationally intractable. Planning-Scheduling papers assume behavior is static. Scheduling-Behavior papers treat infrastructure as given. Planning-Behavior papers treat grid constraints as abstract. None of them are wrong, exactly. None of them are complete.

The behavioral simplification is where the risk concentrates.

Charging stations last twenty years. The models used to site them — by operators, by municipalities, by federal programs like NEVI — often represent driver behavior as a static aggregate: average charging duration, predictable arrival patterns, network effects that average out heterogeneity. But real driver behavior is lumpy, socially contingent, and responsive to exactly the kind of infrastructure the models assume. A station placed assuming drivers will distribute themselves evenly across available ports will be overbuilt in some neighborhoods and a dead zone in others. A grid constraint modeled as a static ceiling is actually a moving target as EV adoption ramps.

This is not theoretical. Grid constraints are already the binding factor for the industry.

More than 90 percent of EV charging operators expect grid capacity to constrain their operations, according to Andrew Bennett, CEO of charging software company Driivz, speaking to GovTech in May 2025. The halt to NEVI federal funding that month — a $5 billion Formula Program — changed almost nothing about what any charge point operator in the United States is actually doing. Not because the money was irrelevant, but because the money was never the bottleneck. The bottleneck is getting power to the charger once it is sited.

If the binding constraint is megawatts rather than capital, the PSB trilemma becomes a capacity problem rather than a funding problem — and the solutions look different. You cannot spend your way out of a trilemma. You can only choose which fidelity sacrifice you are willing to live with.

The academic response to this, the paper suggests, is machine learning. Several research teams are attempting to use ML approaches — the paper cites EV-Planner, a machine learning tool for charging station placement published in Frontiers in AI in early 2026 — to escape the fidelity-tractability tradeoff by learning behavioral patterns from data rather than optimizing against assumed ones. Whether these tools survive contact with the operational realities of grid constraints, charger uptime, and driver churn is an open question. They exist in the research literature. They have not yet displaced the static models in operational planning at most charge point operators.

The infrastructure being built under the old assumptions will outlast the research that justified it.

This is the second-order risk the paper makes visible if it does not prove: that we are making permanent infrastructure decisions using models whose behavioral simplifications are not known to be approximately correct, only known to be computationally necessary. Whether those simplifications are close enough to work — or whether we are systematically misplacing chargers in ways that will require remediation for decades — is not a question the trilemma answers. It is the question whose answer the trilemma makes urgent to go look for.

The William & Mary team has given the field a vocabulary for a problem practitioners already knew existed. What happens next is not a research question. It is a siting decision, being made this quarter, in every city where a charger is going in.

Source: arXiv CS.MA, Planning Scheduling and Behavior in EV Charging Systems: A Critical Survey and Trilemma Framework, Xiao et al., submitted May 20, 2026. GovTech, May 2025. Frontiers in AI, EV-Planner, 2026.