Electric taxi and robotaxi fleets are planned by one company and the local power grid is operated by another, and neither shares its internal model with the other. When a fleet's dispatch software sends dozens of cars to the same fast-charging hub, it can push a transformer past its rating or pull voltages below legal limits at the end of a feeder (a feeder is a distribution line serving a neighborhood). The usual fixes either merge both systems into one giant optimization, which neither party will allow and which does not scale, or have the grid broadcast prices, which drivers and fleets chase in herds. We propose a middle path. At every control step, the grid operator sends the fleet a power envelope: for each charging hub and each upcoming time slot, a minimum and maximum aggregate power that the grid can safely absorb, given its current state and its renewable forecasts. The fleet then re-plans its routing and charging inside those bounds. Both sides use receding horizon control, meaning each plans a short window ahead, applies only the first step, observes what actually happened, and re-plans.
The point that makes this more than plumbing is that a hub's real power draw is not something the fleet sets freely. Fast chargers switch from a constant-current regime, where power is roughly flat, to a constant-voltage regime, where power tapers as the battery fills. This switch is triggered when state of charge crosses a threshold. Systems like this, with smooth evolution punctuated by state-triggered switches, are called hybrid systems. A fleet planner that ignores the taper will promise the grid a load profile it cannot deliver, in either direction. We claim that modeling the taper as a piecewise affine constraint inside the fleet's controller is what makes envelope coordination both safe and efficient in closed loop. We further claim that a bounded, privacy-preserving envelope exchange can recover most of the benefit of a fully joint controller, while prices and static per-hub caps cannot.
Core Claim
Three separate claims can be tested and can fail. First, in closed-loop simulation on a feeder with variable solar and wind, envelope-coupled receding horizon control keeps voltage and transformer violations near zero while losing little passenger service relative to a centralized controller that sees both models. Second, it beats price signals and static per-hub capacity caps on the combined service–safety tradeoff. Third, removing the constant-voltage taper from the fleet model is what breaks it: envelopes that are respected on paper are violated in practice, or are left mostly unused. If the third claim fails, the hybrid modeling is decorative and the contribution reduces to a coordination protocol. We would report that outcome as such.
Positioning Against Prior Work
The fleet side builds directly on Data-Driven Model Predictive Control of Autonomous Mobility-on-Demand Systems. That work re-solves a mixed-integer rebalancing program over a time-expanded network using demand forecasts from a recurrent neural network. It cut mean wait from 283.3 to 29.4 seconds on Hangzhou ride-hailing data, but its vehicles have no batteries and no grid. Joint Optimization of Autonomous Electric Vehicle Fleet Operations and Charging Station Siting adds state of charge and charging modes. It reduces peak charging load by 10.07% in San Francisco, but it is an offline full-day linear program, and it sees the grid only through tariffs and demand charges. Our proposal takes its battery-aware state space, puts it in a rolling loop, and replaces tariffs with physical envelopes.
Two works motivate the claim that coordination is needed and that it is hard. An Open Source Modeling Framework for Interdependent Energy-Transportation-Communication Infrastructure in Smart and Connected Communities shows that ignoring coupling biases peak power predictions by 7% and road speeds by 10.5%, but it is a forward simulator with no controller. Integrated Charging Scheduling and Operational Control for an Electric Bus Network jointly schedules buses and chargers with Lagrangian decomposition and saves 11–19% of operating cost. It treats the grid as an unconstrained source with exogenous prices, assumes a single shared terminal, and runs open loop. That paper also supplies a constraint on our design: Gurobi found no feasible solution within 30 minutes on its 20-line instance. A monolithic joint mixed-integer controller is therefore not a viable operational design, and we use it only as an offline reference.
Price-based coordination is the main conceptual competitor. Optimal Pricing to Manage Electric Vehicles in Coupled Power and Transportation Networks and A multi-objective spatio-temporal pricing method for fast-charging stations oriented to transformer load balancing set charging prices to steer route choices toward balanced transformer loads. Smart Routing of Electric Vehicles for Load Balancing in Smart Grids shows that decentralized routing games approach load-balancing equilibria as fleets grow. These results describe equilibria of many selfish drivers. Our setting is a single managed fleet that reacts strategically and immediately to a price, which is where herding and rebound peaks appear. We test that directly rather than assume it.
The hybrid-control machinery comes from Low-Complexity Method for Hybrid MPC with Local Guarantees. It solves receding horizon problems over piecewise affine dynamics by operator splitting, reducing each iteration to projections onto convex polyhedra. It reaches millisecond runtimes on small mechanical benchmarks, with convergence guaranteed only to local minima. We reuse it for the per-vehicle charging subproblems, where the taper is exactly a union of polyhedra. Its local-only guarantee is a real constraint on how much we can claim, as discussed below.
Finally, Safe and Sustainable Electric Bus Charging Scheduling with Constrained Hierarchical DRL trains a hierarchical constrained reinforcement learning policy for e-bus charging under solar, price, and traffic uncertainty. It comes within about 1% of an oracle on cost, but about 1% of episodes still violate battery constraints, and it does not model feeder power flow. We treat it as the learned-policy competitor.
Research Program
The grid side models one or more radial feeders with the linearized distribution power flow equations known as LinDistFlow. In this model, bus voltages and line flows are affine in nodal injections, so the set of safe hub loads at each time slot is a polytope. Each control step, the grid operator solves a small linear program over its horizon. That program uses current measurements and a set of renewable scenarios, and it returns per-hub, per-slot bounds $[\underline{P}_{h,k}, \overline{P}_{h,k}]$ that are jointly safe across a chosen fraction of scenarios. Here $h$ indexes hubs and $k$ indexes slots in the horizon. Because the true safe set couples hubs that share a feeder, reducing it to independent per-hub boxes loses some flexibility. We will study three ways to split it: proportional to recent use, proportional to the fleet's reported flexibility, and an iterative scheme in which the fleet returns its marginal value for extra headroom at each hub.
The fleet side extends the time-expanded rebalancing program of Iglesias et al. with state-of-charge bins and charging decisions, following Luke et al. The fleet must keep each hub's total draw $\sum_{i \in h} p_{i,k}$ inside the envelope. Each charging vehicle $i$ obeys the taper as a piecewise affine bound,
$$p_{i,k} \le \min\big(P^{\max}_i,\; \alpha_i (1 - s_{i,k})\big),$$
where $p_{i,k}$ is the vehicle's charging power in slot $k$, $s_{i,k}$ is its state of charge as a fraction, $P^{\max}_i$ is the constant-current power limit, and $\alpha_i$ sets the constant-voltage slope. The kink where the two terms meet is the hybrid mode switch. To keep each control step tractable, we split the problem in two. The fleet-level flow problem stays a mixed-integer program over aggregate bins, solved with a time limit. Per-hub charging allocations are then refined by the operator-splitting method applied to the piecewise affine constraints. Whatever the solvers return, a closed-loop simulator replays the actual charging physics, including the taper, and feeds the measured hub loads back to the grid model, which runs a nonlinear power-flow check. Constraint violations are counted in this simulator, not in the controller's own model.
The analysis has three parts. First, we will measure the price of decentralization: the service gap between envelope coordination and a centralized controller on the same instances, as a function of horizon length and of how the shared feeder budget is split. Second, we will measure the cost of ignoring hybrid dynamics: realized minus planned hub load, envelope utilization, and violations, with and without the taper. Third, we will measure how much information each scheme exchanges per step.
Baselines
The naive reference is uncoordinated reactive operation. The fleet runs the Iglesias-style rebalancing controller without any grid information, and vehicles charge at full power whenever state of charge falls below a threshold. It runs on the same demand traces and feeder. It sets the floor for grid violations and the ceiling for unconstrained service. Iglesias et al. report a 29.4-second mean wait for their predictive controller on Hangzhou data, but that is with no batteries. With batteries we expect it to be higher, and we will report our own reimplemented figure rather than borrow theirs.
The strongest existing method for this question is price-signal coordination. The grid operator solves for per-hub, per-slot prices in the spirit of Alizadeh et al. and Guo et al., and these prices enter the fleet controller's objective. We give the grid operator the same LinDistFlow model, the same scenarios, and the same per-step compute as in the envelope scheme. We tune price update rules on held-out days with the same hyperparameter budget. Beating it means fewer violation-minutes at equal or better mean wait, or equal violations at a strictly lower wait, across a full week of test days. There is no reported result on Lacuna for closed-loop voltage violations under price-coordinated fleet charging, so this comparison has no prior number to anchor it. For context, the closest reported coordination figures come from offline or open-loop settings: a 10.07% peak load reduction (Luke et al.) and an 11–19% operational cost reduction (Lacombe et al.).
The static-cap baseline gives each hub a fixed power limit, computed offline from the worst-case feeder hosting capacity, and the fleet controller enforces it. This is what a utility might do today. The centralized controller solves fleet and LinDistFlow together with full information, using the same decomposition plus a generous time limit. It is an upper bound on what coordination can achieve, not a deployable method. The learned-policy baseline is a constrained hierarchical reinforcement learning agent in the style of Qi et al. It is trained on the training days with the feeder in the loop and an envelope-violation penalty, using a compute budget matched to our total offline tuning.
The key ablation is envelope coordination with the taper removed, so the fleet assumes chargers deliver $P^{\max}_i$ until full. It isolates whether hybrid modeling matters. A second ablation replaces the scenario-based envelopes with envelopes computed from a single mean forecast, which isolates the value of uncertainty handling.
Why not just…?
Why not just broadcast prices and let the fleet optimize? Prices are a familiar and privacy-preserving way to coordinate. Etesami et al. show that decentralized routing converges to balanced load as fleets grow. However, a single managed fleet that re-optimizes against a price does not behave like many small drivers. It moves its whole flexible load to the cheapest slot and hub, producing a new peak there. Prices also give no hard guarantee: a price that is slightly too low at one step still overloads a transformer. The price baseline above settles this empirically. If tuned prices match envelopes on violations at equal wait time, the protocol claim fails, and we will say so.
Why not just put a fixed power cap on each hub? This is the most serious threat, because it requires no communication at all. It wins whenever the binding grid limit is a transformer rating that does not depend on weather. It should lose on feeders where the limit is a voltage bound at the far end, because there safe headroom moves with solar output and with other loads. A cap safe in the worst hour then wastes capacity most of the day. We will include a transformer-limited feeder configuration where we expect caps to tie, and a voltage-limited, high-renewable configuration where we expect them to lose. If caps tie in both, the right conclusion is that utilities should publish time-varying caps and skip online exchange.
Why not just merge both models into one joint controller? Both operators guard their network data, and the joint problem is the one Lacombe et al. found Gurobi could not even make feasible within 30 minutes at modest scale. We still run the merged controller offline as an upper bound. A small gap between it and the envelope scheme is the evidence that decentralization costs little.
Why not just train a reinforcement learning policy? Qi et al. reach near-oracle cost, but they retain roughly 1% battery-constraint violations, and they model no feeder. A learned policy also has to be retrained whenever the grid operator changes its envelopes or the feeder changes. Our matched-budget comparison will report violations under held-out weather, the regime where we expect the learned policy to degrade.
Proposed Experiments
The testbed couples fleet demand to feeder physics. For demand, we use the Hangzhou ride-hailing trips across 66 regions used by Iglesias et al., and the San Francisco 190-zone demand used by Luke et al. Charging hubs are mapped onto buses of the IEEE-33 feeder. Renewable scenarios at those buses come from the A Large-Scale Dataset of Distributed Renewable Energy Scenarios on the IEEE-33 Bus Network collection, which provides 300 joint wind–solar scenarios. Scenarios are split into tuning and test sets so that envelope construction never sees test weather. For the multi-hub case, several feeder copies are attached to a shared substation, so that hubs compete for shared headroom. We vary fleet size, charger power, and solar penetration.
Metrics are as follows. Service is measured by mean and 95th-percentile passenger wait. Grid safety is measured by minutes outside voltage limits and transformer overload energy, both computed by nonlinear power flow in the simulator. Cost covers energy and the cost of empty rebalancing miles. Controller behavior is measured by per-step solve time against a 60-second control period, and by the number of values exchanged per step.
The first decisive check is the taper ablation. If removing the constant-voltage model leaves both violations and service unchanged, the hybrid component is not needed. The second is the static-cap comparison on the voltage-limited, high-renewable feeder, which decides whether online exchange is worth having. The third is the gap to the centralized controller. We expect it to be small for horizons of one to two hours, and we will report how it grows with the share of hubs on a shared feeder. That growth tells us when splitting the feeder budget into per-hub boxes stops being enough.
We will also report how often the operator-splitting refinement stalls at a poor local solution, because the method's guarantees are only local. We will measure this by comparing its output with a time-limited mixed-integer solve on a subsample of steps. A high stall rate would mean the scalability claim rests on the decomposition and not on the hybrid solver.
References
- Data-Driven Model Predictive Control of Autonomous Mobility-on-Demand Systems
- Joint Optimization of Autonomous Electric Vehicle Fleet Operations and Charging Station Siting
- An Open Source Modeling Framework for Interdependent Energy-Transportation-Communication Infrastructure in Smart and Connected Communities
- Integrated charging scheduling and operational control for an electric bus network
- Optimal Pricing to Manage Electric Vehicles in Coupled Power and Transportation Networks
- A multi-objective spatio-temporal pricing method for fast-charging stations oriented to transformer load balancing
- Smart routing of electric vehicles for load balancing in smart grids
- Low-Complexity Method for Hybrid MPC with Local Guarantees
- Safe and Sustainable Electric Bus Charging Scheduling With Constrained Hierarchical DRL
- A Large-Scale Dataset of Distributed Renewable Energy Scenarios on the IEEE-33 Bus Network