Zonal Dispatching Control

• Zonal Dispatching Control is a paradigm that partitions large systems into spatial or electrical zones to reduce control complexity and optimize localized decision-making.
• It employs methods like sensitivity-based clustering, MPC, and RL to determine optimal zone boundaries and manage intra- and inter-zone interactions.
• Hierarchical and multi-timescale strategies integrate coarse global planning with fast local adjustments, yielding improved service quality and reduced computational burdens.

Zonal dispatching control denotes a class of control architectures in which a large networked system is partitioned into spatial or electrical zones, and dispatch decisions are taken at the zone level rather than only at the level of individual devices, vehicles, or loads. In the cited literature, the zone may be a weakly coupled voltage-control area on an unbalanced feeder, a taxi zone, a virtual AMoD region, a charging-station catchment, a flexible-route segment, a traction station cluster, or an entire active distribution network behind a grid connection point (Alrushoud et al., 2021, Riley et al., 2020, Song et al., 2022, Gupta et al., 2024, Li et al., 2024). Across these formulations, zonal dispatching control combines aggregation, local feasibility, and reduced-dimensional decision-making: it replaces a single global real-time control problem with a collection of local or regional control problems, often coordinated hierarchically across timescales.

1. Zonal abstraction and system boundaries

The defining feature of zonal dispatching control is the replacement of a fully nodal or continuous-space description by a partition into operational subdomains. In distribution-system Volt/VAR control, the feeder is partitioned into non-overlapping, weakly coupled voltage-control zones, typically per phase, so that each zone can be treated as a quasi-independent local area for voltage regulation (Alrushoud et al., 2021, Alrushoud et al., 2020). In ride-sharing and mobility-on-demand, zones are taxi zones, census tracts, virtual regions, or graph nodes, and dispatching acts on counts of vehicles moving between zones rather than on raw road-network trajectories (Riley et al., 2020, Aalipour et al., 2024, Zhu et al., 2024). In EV charging dispatch, each charging station induces an implicit station-based region described by distance, route congestion, and queue length (Song et al., 2022). In traction power, each traction station cluster is an electrically decoupled local zone with its own phase-angle controller (Li et al., 2024). In active distribution networks, the entire ADN behind the grid connection point is treated as a single dispatchable zone whose net power at the GCP is the dispatch variable (Gupta et al., 2024, Gupta et al., 2024).

Setting	Zone definition	Primary dispatch variable
Distribution Volt/VAR	Weakly coupled voltage-control zone	IBR reactive power, VR setpoint (Alrushoud et al., 2021)
Ride-sharing	Taxi zone / OD zone	Idle-vehicle relocation counts (Riley et al., 2020)
AMoD	Virtual region or zone node	Occupied and empty inter-zone flows (Aalipour et al., 2024)
EV charging dispatch	Station-centered implicit zone	Charging-station choice (Song et al., 2022)
Traction power	Traction station cluster	A-TS phase angle (Li et al., 2024)
Active distribution network	Entire feeder behind GCP	Net active power at GCP (Gupta et al., 2024)

This suggests that “zone” is not a single geometric construct but a modeling interface. A zone is chosen so that internal actuation remains meaningful and external coupling remains manageable. In some settings, such as the wind-farm dispatch problem, the literature explicitly interprets the whole farm as a single zone whose internal units track a zonal reference with fair sharing (Baros, 2016).

2. Zone formation and decomposition principles

Zone formation is domain-specific, but the recurrent principle is strong intra-zone coupling and weaker inter-zone coupling. In unbalanced distribution feeders, zoning is derived from a voltage–reactive-power sensitivity matrix. For each phase, repeated power-flow perturbations define

$$q_{i,j}=\frac{V_{ij}-V_i^{\text{BASE}}}{\Delta Q_j},$$

and Pearson correlation between the columns of $\mathrm{VLSM}_Q$ gives a similarity measure for clustering (Alrushoud et al., 2021, Alrushoud et al., 2020). Fast incremental clustering then assigns node $l$ to zone $k$ if its mean correlation

$$\mathrm{mcc}_{l,k}=\frac{1}{m_k}\sum_{v\in k} C_{l,v}$$

exceeds a threshold $\alpha$; otherwise a new zone is created. In the 1283-node feeder case, $\alpha=0.96$ yielded 5 IBR zones and 4 VR zones (Alrushoud et al., 2021).

In dynamic ride-hailing relocation, zone formation can itself be time-varying. The dFDA-VeD framework partitions the road graph into subareas and relocation centers by minimizing

$$F(\mathcal{C},V,W_v)=\sum_{i=1}^n \min_{c_j\in\mathcal{C}} d(c_j,v_i)\,S(g_i),$$

where $g_i^t$ is the pickup–dropoff gap at vertex $v_i$ and $\mathrm{VLSM}_Q$0 weights undersupplied points (Guo et al., 2021). This yields dynamic, demand-weighted zones rather than fixed administrative ones.

Other systems adopt coarser or exogenous partitions. The AMoD discrete-time control model uses a city partitioned into virtual zones and a transformed complete directed graph over those zones (Aalipour et al., 2024). The Shenzhen hierarchical repositioning controller first partitions the city into $\mathrm{VLSM}_Q$1 non-overlapping regions using K-means clustering (Zhu et al., 2024). The Munich semi-on-demand feeder fixes a 1.2 km fixed-route segment followed by two 2.2 km flexible zones (Ng et al., 2 Sep 2025). The meal-delivery framework uses a fixed $\mathrm{VLSM}_Q$2 set of H3 hexagonal cells as operational zones (Cheng et al., 10 Jan 2025). The EV charging framework does not explicitly name zones, but its station-indexed state variables define an implicit partition into charging-station regions (Song et al., 2022).

These constructions reveal two distinct zoning logics. One is endogenous and physics- or demand-derived, as in sensitivity clustering or demand-weighted medoids. The other is exogenous and planning-driven, as in predefined transit zones, H3 grids, or traction station clusters. The former emphasizes control decoupling; the latter emphasizes tractability and operational regularity.

3. Hierarchical and multi-timescale control structures

A large fraction of zonal dispatching architectures are hierarchical. In the two-stage zonal Volt/VAR controller, fast inverter-based resources dispatch reactive power every 1 minute within IBR zones, while voltage regulators receive setpoint updates every 2 minutes in overlapping VR zones; the separation is enforced both by different voltage bands and by VR time delays of $\mathrm{VLSM}_Q$3–$\mathrm{VLSM}_Q$4 s (Alrushoud et al., 2021). In large-scale ride-sharing, request-level dispatching runs at epochs of length $\mathrm{VLSM}_Q$5 such as $\mathrm{VLSM}_Q$6 s, while zone-level relocation uses MPC at periods $\mathrm{VLSM}_Q$7 such as $\mathrm{VLSM}_Q$8 s (Riley et al., 2020). In the ADN dispatch framework, the day-ahead layer sets a 5-minute dispatch plan at the GCP, while intra-day MPC actuates every 30 s to track it (Gupta et al., 2024, Gupta et al., 2024). In the Shenzhen AMoD controller, DeePC computes inter-regional flows every $\mathrm{VLSM}_Q$9 minutes, while coverage control updates vehicle targets every $l$0 s (Zhu et al., 2024). In the semi-on-demand SAV feeder, RL assigns vehicles to route patterns or flexible zones at terminus departure epochs, while route insertion remains a lower-level deterministic schedule optimization (Ng et al., 2 Sep 2025).

This hierarchy usually separates a slower, aggregated balancing problem from a faster local actuation problem. The control signal at the upper layer is typically low-dimensional: zonal relocation counts $l$1 in ride-sharing (Riley et al., 2020), inter-regional flows $l$2 in DeePC-based AMoD (Zhu et al., 2024), a dispatch plan $l$3 at the GCP (Gupta et al., 2024), or a single A-TS phase angle $l$4 in a traction station cluster (Li et al., 2024). The lower layer then realizes that signal through route selection, voltage setpoints, local charging power, or motion on the road graph.

The hierarchical structure is not always optimization-based. Some systems rely on rule-based zonal control: the distribution-system Volt/VAR scheme dispatches IBRs through a priority list and tunes VR setpoints without retrofitting existing controllers (Alrushoud et al., 2021). Others embed the hierarchy inside learning: the meal-delivery framework iteratively trains dispatching and steering policies; the transit feeder uses PPO at the zonal assignment layer and deterministic insertion at the route layer (Cheng et al., 10 Jan 2025, Ng et al., 2 Sep 2025).

4. Mathematical formulations of zonal dispatch

Three mathematical templates recur.

First, there are sensitivity-based formulations in network control. In zonal Volt/VAR control, the required reactive support for the extreme critical node $l$5 in zone $l$6 is approximated by

$l$7

and inverter dispatch follows a zonal priority list built from correlations with the extreme node (Alrushoud et al., 2021). In traction power, local TS complex powers are explicit functions of the single zonal phase variable $l$8, so power-distribution or constant-power modes reduce to solving one scalar nonlinear equation in $l$9 under a feasible phase-angle domain $k$0 (Li et al., 2024).

Second, there are flow-balance formulations on zonal graphs. The ride-sharing MPC minimizes discounted unserved demand and relocation cost,

$k$1

subject to demand balance and vehicle conservation across zones (Riley et al., 2020). The AMoD discrete-time model distinguishes customer queues $k$2, idle vehicles $k$3, and vehicles in transit $k$4, with controls $k$5 for occupied departures and $k$6 for empty rebalancing; MPC then tracks a rebalancing equilibrium while penalizing waiting queues and long empty moves (Aalipour et al., 2024). The DeePC regional controller similarly uses zone-to-zone flows $k$7 and served-demand outputs $k$8, but replaces an explicit state-space model by Hankel-matrix trajectory constraints learned from data (Zhu et al., 2024).

Third, there are MDP and RL formulations. The EV charging dispatcher defines state $k$9, where $$\mathrm{mcc}_{l,k}=\frac{1}{m_k}\sum_{v\in k} C_{l,v}$$0, $$\mathrm{mcc}_{l,k}=\frac{1}{m_k}\sum_{v\in k} C_{l,v}$$1, and $$\mathrm{mcc}_{l,k}=\frac{1}{m_k}\sum_{v\in k} C_{l,v}$$2 are station-indexed distance, road-congestion, and station-usage vectors, and action $$\mathrm{mcc}_{l,k}=\frac{1}{m_k}\sum_{v\in k} C_{l,v}$$3 selects the charging station (Song et al., 2022). The Munich feeder treats zonal dispatch as an MDP with a compact 18-feature state, action set $$\mathrm{mcc}_{l,k}=\frac{1}{m_k}\sum_{v\in k} C_{l,v}$$4 for regular route, Zone 1, Zone 2, or hold, and reward equal to the negative number of request rejections due to prolonged waiting (Ng et al., 2 Sep 2025). The meal-delivery platform models order assignment and idle-courier steering as separate MDPs over H3 cells, with supply–demand forecasts embedded in the state (Cheng et al., 10 Jan 2025).

These formulations differ in algorithmic machinery—rule-based dispatch, MILP, MPC, DeePC, DQN, dueling DDQN, PPO—but they share a reduced control space induced by zoning. The reduction is not merely computational. It encodes which couplings are considered essential and which are intentionally suppressed.

5. Forecasting, measurement, and information reduction

Zonal dispatching control is often paired with aggressive state compression. In the two-stage distribution controller, critical-node identification by Monte Carlo quasi-static simulation reduced real-time measurement needs to 51 distinct critical nodes on a 1283-node feeder, about 3.9% of nodes (Alrushoud et al., 2021). The controller then uses voltage measurements only from those nodes. In the active distribution network framework, short-term forecasts of demand and PV are refreshed every 30 s, while PMUs and real-time state estimation update the linearized grid model used by MPC (Gupta et al., 2024). The day-ahead layer builds EV scenarios from Gaussian-mixture models over arrival, departure, SoC target, battery capacity, and power limits, and combines them with load and PV scenarios in a scenario-based stochastic OPF (Gupta et al., 2024).

Mobility formulations also rely on zonal prediction. The ride-sharing controller predicts per-zone aggregate demand using weekly differencing and VAR across adjacent zones, then maps origins to destinations using historical OD proportions to obtain $$\mathrm{mcc}_{l,k}=\frac{1}{m_k}\sum_{v\in k} C_{l,v}$$5 (Riley et al., 2020). The dFDA-VeD system uses historical averages to predict point-level pickups and dropoffs, then aggregates them into zonal demand–supply gaps that drive dynamic partitioning (Guo et al., 2021). The meal-delivery dual-control framework explicitly injects predicted demand into both dispatching and steering states, while the steering policy uses a mean-field neighborhood score to summarize local zonal imbalance (Cheng et al., 10 Jan 2025). The transit feeder includes a 15-minute demand forecast in the PPO state (Ng et al., 2 Sep 2025).

This information compression does not eliminate uncertainty; it relocates it. Prediction error becomes a first-order issue once the system is controlled through zonal aggregates rather than raw states. Several papers make that dependence explicit. Ride-sharing gains are reported to be strongest in high-demand instances and more sensitive to prediction quality in sparse zones (Riley et al., 2020). The ADN MPC is deterministic and relies on rolling re-optimization every 30 s to absorb forecast mismatch (Gupta et al., 2024).

6. Empirical effects, trade-offs, and limits

The empirical literature consistently reports that zonal dispatching control improves at least one of service quality, mechanical wear, or computational tractability, but not without trade-offs.

In distribution voltage control, the two-stage zonal method maintained all nodes within ANSI limits and reduced VR tap operations, while using only a small critical-node set and no VR retrofits (Alrushoud et al., 2021). In the EPFL ADN, tracking the dispatch plan with MPC reduced Day 1 error metrics from RMSE $$\mathrm{mcc}_{l,k}=\frac{1}{m_k}\sum_{v\in k} C_{l,v}$$6 kW, AEE $$\mathrm{mcc}_{l,k}=\frac{1}{m_k}\sum_{v\in k} C_{l,v}$$7 kWh, and MAE $$\mathrm{mcc}_{l,k}=\frac{1}{m_k}\sum_{v\in k} C_{l,v}$$8 kW without control to RMSE $$\mathrm{mcc}_{l,k}=\frac{1}{m_k}\sum_{v\in k} C_{l,v}$$9 kW, AEE $\alpha$0 kWh, and MAE $\alpha$1 kW with MPC; Day 2 similarly dropped from RMSE $\alpha$2 kW, AEE $\alpha$3 kWh, and MAE $\alpha$4 kW to RMSE $\alpha$5 kW, AEE $\alpha$6 kWh, and MAE $\alpha$7 kW (Gupta et al., 2024). The corresponding day-ahead/real-time study reported net aggregated energy error reductions of $\alpha$8–$\alpha$9, MAE reductions of $\alpha=0.96$0–$\alpha=0.96$1, and maximum GCP power excursion reductions of $\alpha=0.96$2–$\alpha=0.96$3 when EVCS and BESS flexibility were used jointly (Gupta et al., 2024).

In ride-sharing, integrating zone-level MPC relocation with request-level dispatch reduced average waiting time from $\alpha=0.96$4 minutes to $\alpha=0.96$5 minutes, about a 30% reduction, with standard deviation dropping from $\alpha=0.96$6 to $\alpha=0.96$7, and the largest zonal improvement reaching almost 55% in high-demand zones (Riley et al., 2020). In graph-based AMoD MPC, the LMPC$\alpha=0.96$8 variant achieved average waiting time $\alpha=0.96$9 minutes and total empty distance $$F(\mathcal{C},V,W_v)=\sum_{i=1}^n \min_{c_j\in\mathcal{C}} d(c_j,v_i)\,S(g_i),$$0 miles in one case study, versus $$F(\mathcal{C},V,W_v)=\sum_{i=1}^n \min_{c_j\in\mathcal{C}} d(c_j,v_i)\,S(g_i),$$1 minutes and $$F(\mathcal{C},V,W_v)=\sum_{i=1}^n \min_{c_j\in\mathcal{C}} d(c_j,v_i)\,S(g_i),$$2 miles for IARR (Aalipour et al., 2024). In Shenzhen, the hierarchical “Upper + Lower” controller reached answer rate $$F(\mathcal{C},V,W_v)=\sum_{i=1}^n \min_{c_j\in\mathcal{C}} d(c_j,v_i)\,S(g_i),$$3, waiting time $$F(\mathcal{C},V,W_v)=\sum_{i=1}^n \min_{c_j\in\mathcal{C}} d(c_j,v_i)\,S(g_i),$$4 s, rebalancing distance $$F(\mathcal{C},V,W_v)=\sum_{i=1}^n \min_{c_j\in\mathcal{C}} d(c_j,v_i)\,S(g_i),$$5 km, and VUR $$F(\mathcal{C},V,W_v)=\sum_{i=1}^n \min_{c_j\in\mathcal{C}} d(c_j,v_i)\,S(g_i),$$6, outperforming “No Control” at $$F(\mathcal{C},V,W_v)=\sum_{i=1}^n \min_{c_j\in\mathcal{C}} d(c_j,v_i)\,S(g_i),$$7, $$F(\mathcal{C},V,W_v)=\sum_{i=1}^n \min_{c_j\in\mathcal{C}} d(c_j,v_i)\,S(g_i),$$8 s, $$F(\mathcal{C},V,W_v)=\sum_{i=1}^n \min_{c_j\in\mathcal{C}} d(c_j,v_i)\,S(g_i),$$9 km, and $g_i^t$0 (Zhu et al., 2024). The dynamic demand-aware dFDA-VeD method increased served-trip ratio and utilization while keeping the largest increase in deadheading cost $g_i^t$1 below 5% and waiting-time differences below 4 seconds (Guo et al., 2021).

In EV charging dispatch, Dueling DDQN reduced average travel time to $g_i^t$2 seconds in the 400-EV scenario, 39% less than DQN in that scenario (Song et al., 2022). In semi-on-demand transit, dynamic zonal control served 16% more passengers at 13% higher generalized costs on average compared to traditional fixed-route service, and the RL controller added 2.4% more passengers at 1.4% higher costs relative to nominal zonal dispatch (Ng et al., 2 Sep 2025).

A common misconception is that tighter zoning is always superior. The energy-constrained spatial-queueing study shows the opposite can hold: pooling is preferable under light demand, whereas tighter zoning becomes more profitable as demand increases (Lin et al., 25 Jun 2026). A second misconception is that battery expansion is the dominant lever in energy-constrained zonal systems. Over the tested range, profitability was driven more by zoning than by battery improvement, and larger batteries could delay replenishment and reduce fleet readiness under sparse demand (Lin et al., 25 Jun 2026).

The main limitations are also recurrent. Zonal rules are often heuristic rather than globally optimal (Alrushoud et al., 2021). Performance can be sensitive to forecast quality, especially in sparse zones (Riley et al., 2020). Static zone definitions may become suboptimal under topology changes or demand regime shifts (Guo et al., 2021, Ng et al., 2 Sep 2025). Increasing the number of zones enlarges the control dimension in regional MPC and DeePC, and may erode the very tractability that motivates zoning (Zhu et al., 2024). These limitations do not negate zonal dispatching control; they define its operating envelope. The literature therefore increasingly treats zoning not as a fixed preprocessing step, but as a design variable coupled to forecasting, control architecture, and physical feasibility.