Arrival-Time Pricing for Solar-Fed EV Charging Hubs on Real Road Networks

When an electric vehicle sets off for a charging hub, the driver commits to a choice now that pays off twenty or forty minutes later. If that hub is powered by rooftop or canopy solar, the charging capacity the driver will actually find depends on the clouds at arrival, not at departure. Selfish drivers who all react to today's posted availability can pile onto the same hub, then find it throttled by a passing weather front. The result is that the gap between selfish and coordinated outcomes is set by a race between how fast renewable supply changes and how long drivers stay locked into their decisions. That gap is usually measured by the price of anarchy, the ratio of total cost when everyone acts selfishly to total cost under the best coordinated plan.

We propose to study and price that race directly. The core observation is structural. On any real road network, the final decision of which hub to charge at is a choice among parallel options, even when the roads leading there form a tangled mesh. Parallel-option games are the setting where equilibria and efficient tolls have closed forms. So we can price hubs analytically, using a forecast of each hub's solar capacity at the driver's expected arrival time, and recompute those prices every time the forecast updates. The slower, mesh-wide road equilibrium is left to a standard solver. Our claim is that this arrival-time hub pricing recovers most of the efficiency that intermittency destroys. We also predict exactly when it fails: when hubs share congested road bottlenecks, so that the parallel structure is only an approximation.

Where this sits in prior work

The closest work is Green Routing Game. It models logistics fleets of combustion and electric vehicles routing over parallel paths. Charging delay is zero up to a fixed "cost-free" capacity and convex beyond it. The paper reports a price of anarchy "as high as 187.6" without pollution taxes and "just 1.2" under a calibrated tax. That paper motivates ours, and it also leaves three things open. Its topology is strictly parallel, and the authors name urban grids as future work. Its charging capacity is static and detached from any power supply. Its tax is fixed, so it cannot respond when capacity moves.

On mesh networks, Interpolated Traffic Assignment (I-TAP) blends selfish and system-optimal routing with a single parameter. It stays convex and solves with Frank-Wolfe, a standard iterative method for convex traffic assignment. It is fast, reporting 0.03 s on Sioux Falls versus 20 s for constrained approaches. It is also static, with no energy state and no stochastic supply. We reuse it as the road-layer solver and as a strong baseline.

On the power side, Optimal Dispatch of Residential PV Inverters shows how tail-risk accounting changes physical decisions. It uses CVaR, the expected loss in the worst few percent of outcomes, and finds curtailment rises from 27.83 kWh (risk-neutral) to 44.05 kWh (95% confidence) on a radial feeder. However, it treats loads as passive and fully compliant. We borrow its CVaR treatment of forecast error, but apply it to strategic, mobile loads.

Two results constrain the design.

The parallel-link learning mechanisms of Li et al. (Selective Information Disclosure, CHAR) are a different instrument: they control what information drivers see, not what drivers pay. They also share the parallel restriction. We do not compete with them, but the hub layer is a natural place to port them later.

Core Claim

Let $x_h$ be the flow of vehicles arriving at hub $h$. Let $c_h$ be its random charging capacity, driven by local solar output and feeder headroom. Let $w_h(x_h, c_h)$ be the per-vehicle waiting cost. A vehicle departing at time $t$ with expected travel time $\tau$ is charged a hub price equal to its congestion externality, evaluated under the forecast of capacity at arrival:

$$p_h(t) = \mathrm{CVaR}_\beta\big[\, x_h \,\partial w_h(x_h, c)/\partial x_h \,\big], \quad c \sim F_h(t+\tau).$$

Here $F_h(t+\tau)$ is the forecast distribution of hub capacity at arrival, and $\beta$ is the risk level ($\beta \to 0$ recovers the expected-value price). For a fixed road-cost vector, the hub layer is a parallel-link game. That makes $p_h$ a closed-form or one-dimensional root-finding computation per hub.

Our hypotheses are three:

  1. Under realistic intermittency, prices computed from current capacity leave a large share of the price-of-anarchy gap unrecovered.
  2. Arrival-time prices close most of that gap.
  3. The residual gap grows with a measurable coupling index: the share of total road delay incurred on links used by paths to two or more hubs.

On the theory side, we will try to bound the efficiency loss of hub-only pricing by that coupling index times the road cost's sensitivity to flow. Reaching such a bound is plausible because, with zero shared congested links, the game separates exactly into independent parallel subgames.

Research Program

We first build a quasi-dynamic simulator. Each 15-minute period carries departure cohorts that are assigned through a mesh road network to hubs. Cohorts arrive in a later period determined by their equilibrium travel time, and they meet the capacity realized at that later time. This is deliberately simpler than full dynamic traffic assignment. The I-TAP authors leave dynamic assignment open, and we only need the travel-time lag to be faithful.

Hub capacities come from the DDRE-33 dataset. It provides 2,000 single-node wind and PV series and 300 joint scenarios at 15-minute resolution on the IEEE-33 radial feeder, and reports lag-1 autocorrelation above 0.93. Hubs are attached to feeder buses, so hubs on the same branch share headroom.

The theory work has two parts. We characterize when hub-only prices are exactly efficient, and we derive the coupling-index bound. Throughout, we report the realized price of anarchy against a hindsight system optimum that knows true capacities.

Baselines

The naive reference point is unmanaged selfish routing with live broadcast of current hub availability. This is what a charging app does today, and the herding literature cited above predicts it will oscillate. It runs on the same networks, demand, and capacity scenarios, with no tuning.

The strongest existing competitor is myopic full re-solve. Every 15-minute period, we recompute a system-optimal marginal-cost assignment over the whole mesh using current hub capacities as fixed parameters, with the I-TAP Frank-Wolfe solver Interpolated Traffic Assignment at $\alpha = 1$. We also run I-TAP at intermediate $\alpha$ for the fairness comparison. It gets the same demand, scenarios, and wall-clock budget per period, and its convergence tolerance is tuned on held-out scenarios exactly as ours is.

The reported reference numbers come from different, static settings:

  • I-TAP: 0.03 s on Sioux Falls, and a 50% unfairness reduction at about 2% more total travel time.
  • Green Routing Game: price of anarchy 1.2 under a calibrated tax on parallel paths.

There is no reported result on Lacuna for realized price of anarchy under intermittent charging capacity on a mesh. Beating this baseline means a lower realized price of anarchy and lower charging overload at equal or lower compute. We will report the margin with confidence intervals across the 300 joint scenarios.

The second strong competitor is a static tax tuned to mean capacity, as in Green Routing Game, ported to hubs. Its single price per hub is tuned by grid search on training scenarios with the same budget we use to choose $\beta$.

The third is conservative planning, which assigns flows against a fixed low-percentile capacity. This is the CVaR logic of Optimal Dispatch of Residential PV Inverters without strategic updating, and it should trade overloads for wasted capacity. A perfect-foresight system optimum gives the lower bound.

Two ablations isolate our ingredients:

  • Forecast versus current capacity: current-capacity hub prices, identical in every respect except evaluating $F_h$ at $t$ rather than $t+\tau$.
  • Hub layer versus whole mesh: forecast-based marginal-cost tolls on every road link versus hub-only prices.

Why not just…?

Why not just re-run full Frank-Wolfe every period with the new capacities? This is the real threat. At 0.03 s on Sioux Falls, compute is not the binding latency at 15-minute resolution. Our argument is that the binding latency is driver commitment, not solver time. A fresh solve on current capacity still prices the wrong moment. The forecast-versus-current ablation settles this directly. If myopic full re-solve matches arrival-time pricing, the forecasting claim is wrong and we will say so. We will also time full re-solve on the Chengdu network (1,902 nodes, 5,943 links, Urban link travel speed dataset) to check whether compute becomes binding at megacity scale or at sub-period update rates. We do not assume it does.

Why not just a well-calibrated fixed tax? Green Routing Game shows taxes can bring the price of anarchy to 1.2, but with fixed capacity. Its own finding is that operators abandon electric fleets when charging delay spikes, which suggests fixed prices fail exactly at capacity drops. Comparing on low-solar scenarios from DDRE-33 against high-solar ones will show whether the fixed tax's loss concentrates in those episodes.

Why not just plan for the worst-case capacity? It avoids overloads but strands solar energy on good days. That mirrors the curtailment increase from 27.83 to 44.05 kWh in Optimal Dispatch of Residential PV Inverters. We report both overload frequency and unused hub energy, so that this baseline cannot win on one axis by hiding the other.

Why not just apply a parallel-network mechanism after collapsing the mesh into a few routes? This is roughly the test bed of Karma economies, which uses three parallel routes. Collapsing the mesh discards the shared road bottlenecks that our coupling index measures. The hub-layer-versus-whole-mesh ablation, swept across networks with different coupling, shows how much efficiency that discards.

Proposed Experiments

Networks. Road meshes are Sioux Falls, Anaheim, and the three Berlin sub-networks used by I-TAP, plus Chengdu for scale. In each mesh, 4 to 12 hubs are mapped onto DDRE-33 feeder buses.

Metrics.

  • Realized price of anarchy against hindsight optimum.
  • Total travel plus charging delay.
  • Hub overload frequency and unused solar energy.
  • I-TAP's unfairness measure, so equity is not traded silently.
  • Wall-clock time per update.

First check. Does intermittency inflate the realized price of anarchy of myopic full re-solve well above its static value? If not, the problem is not worth a mechanism.

Second check. Does the forecast-versus-current ablation produce a gap that grows with travel time $\tau$ and shrinks with capacity autocorrelation? The model predicts both directions, so either failing falsifies the mechanism.

Third check. As we vary hub placement to raise the coupling index, does the advantage of hub-only pricing over whole-mesh tolls decay roughly as the bound predicts? This tells us where the parallel decomposition is safe.

Microscopic validation. We will replay the best and worst configurations in SUMO on the Manhattan Upper East Side network from AlphaRoute. This checks that the quasi-dynamic lag model does not create or erase the effect, using AlphaRoute's reported 12.88% latency improvement over minimum-distance routing only as a sanity reference for that network.