Power Distribution: Modeling, Optimization, and Control

• Power distribution is the management of medium- and low-voltage electric networks that deliver power from substations to customers through primarily radial configurations.
• Mathematical models such as DistFlow and branch-flow, along with linearizations and convex relaxations, enable efficient analysis and optimal reconfiguration of power flows.
• Current research emphasizes synthetic network generation and topology inference from smart-meter data to support reliable integration of distributed energy resources.

Power distribution, in electric power systems, concerns the medium-/low-voltage infrastructure, mathematical models, and control procedures that connect high-voltage substations to transformers, feeders, and end-use customers, typically through radial or quasi-radial networks. Contemporary research treats it as a coupled problem of topology knowledge, power-flow modeling, voltage regulation, phase balance, distributed optimization, and data realism under high penetrations of distributed energy resources, especially single-phase EV chargers, rooftop PV, and inverter-based devices. The same term also appears in statistics, signal processing, machine learning, and instrumentation, where it denotes specific probability laws, sequence-level sampling targets, or engineered supply architectures (Meyur et al., 2022, Tomihari et al., 6 May 2026).

1. Physical scope in electric power distribution

Low-voltage distribution networks are typically radial or quasi-radial, so that each customer’s meter lies downstream of exactly one phase terminal of a distribution transformer. Accurate knowledge of network topology—including which customers are on which feeder phase—is critical because, without it, operators cannot balance single-phase loads, perform reliable state estimation, fault location or outage management, or integrate distributed energy resources in an optimal fashion (Satya et al., 2016).

Recent work on geographically faithful distribution models represents the network as a hierarchy of substations, primary feeders, pole-top transformers, and residences. In these constructions, road networks, geo-referenced residences, and high-voltage substations are combined to produce medium-/low-voltage systems with nodes labeled as residence, transformer, or substation, and edges labeled as feeder lines, primary lines, or secondary lines. This embedding of electrical infrastructure into road and building geography is used to synthesize networks that resemble their physical counterparts for a given region and to generate ensembles of alternative yet feasible feeder realizations (Meyur et al., 2022).

Operationally, unbalance is a central concern. Unbalanced devices, such as electric vehicle chargers and single-phase solar plants, can exacerbate phase-loading asymmetry, increase losses, trigger protection devices, and potentially damage devices. Utilities therefore perform phase swapping to mitigate overloading and voltage unbalance caused by uneven single-phase injections. This suggests that power distribution is not only a network-construction problem but also a recurrent network-reconfiguration problem tied to evolving load and DER patterns (Gupta et al., 12 Jan 2025).

2. Network equations, linearizations, and relaxation frameworks

A standard analytical basis is the DistFlow or branch-flow formulation on radial feeders. In one simplified radial-feeder model, with branch impedance $r_k+jx_k$, net real/reactive injection $p_k,q_k$, and the homogeneity assumption $x_k/r_k=\alpha$, the linearized recurrences become

$$S_k=S_{k+1}+s_{k+1}, \qquad V_k=V_{k+1}+\rho_k S_k,$$

where $S_k=P_k+\alpha Q_k$, $s_k=p_k+\alpha q_k$, and $\rho_k=r_k/V_0$. This model supports dynamic-programming computation of the distribution of maximal voltage drop under statistically independent random injections, with effectively linear complexity in the number of buses (Turitsyn, 2010).

For radial OPF and control, LinDistFlow and Lin3DistFlow are widely used. In the three-phase phase-allocation setting, the linearized voltage-drop equation is

$$v_{k,\phi}=v_{l,\phi}-2\sum_{\phi'}\left(r_{lk}^{\phi\phi'}p_{lk,\phi'}+x_{lk}^{\phi\phi'}q_{lk,\phi'}\right),$$

with nodal power-balance, injection-linking, phase-count, phase-consistency, and voltage-limit constraints. In radial AC-OPF for uncertain injections, the lossless LinDistFlow balance is written with squared voltages $u_i=v_i^2$, line flows $f_i^P,f_i^Q$, resistances $p_k,q_k$0, and reactances $p_k,q_k$1, yielding affine flow conservation and linearized voltage drops that are compatible with convex chance-constrained reformulations (Gupta et al., 12 Jan 2025, Mieth et al., 2018).

More exact unbalanced formulations use lifted branch-flow variables. In the unbalanced multiphase branch-flow model, one defines $p_k,q_k$2, $p_k,q_k$3, and $p_k,q_k$4, then enforces positive semidefiniteness and, in the exact problem, a rank-one condition. Dropping the rank-one condition yields the convex SDP relaxation. A related voltage-regulation formulation introduces $p_k,q_k$5 and linear trace expressions for injections, flows, and losses; dropping $p_k,q_k$6 gives a convex SDP that is exact under angle-difference and reactive-lower-bound conditions that the paper reports are expected to be satisfied by most networks (Peng et al., 2015, Zhang et al., 2012).

A separate line of work replaces nonlinear OPF by a quadratic approximation derived from a linearized three-phase unbalanced load flow. After expressing voltages as an affine function of generator injections,

$p_k,q_k$7

the loss objective becomes a convex quadratic program in $p_k,q_k$8. For the unconstrained case, the first-order conditions give the analytical solution

$p_k,q_k$9

The reported studies indicate that this approximation is very accurate for systems with a good voltage profile and useful as a warm start for more precise nonlinear formulations (Garces, 2015).

3. Topology inference and synthetic network construction

Topology identification from smart-meter data exploits conservation laws and algebraic structure. In the PCA-based approach for low-voltage systems, one forms the measurement matrix $x_k/r_k=\alpha$0 from energy readings, estimates the mean technical-loss vector, whitens the zero-mean data with the Cholesky factor of the error covariance, computes the sample covariance, and performs eigen-decomposition or SVD to extract the small-variance constraint subspace. Partitioning the estimated constraint matrix into dependent and independent variables yields the regression matrix $x_k/r_k=\alpha$1, which is, up to sign, the incidence submatrix of the feeder graph. Rounding $x_k/r_k=\alpha$2 to a $x_k/r_k=\alpha$3 matrix recovers phase connectivity and, layer by layer, full feeder topology (Satya et al., 2016).

The reported results are exact under the stated sampling conditions. Phase identification achieves 100% accuracy for $x_k/r_k=\alpha$4 samples with computation times on the order of $x_k/r_k=\alpha$5 s. Full-topology identification on the IEEE Roy-Billinton test system achieves 100% accuracy for $x_k/r_k=\alpha$6 samples, with computation time ranging from 4 s for $x_k/r_k=\alpha$7 to 18 s for $x_k/r_k=\alpha$8 (Satya et al., 2016).

Synthetic-network generation addresses the opposite problem: constructing physically plausible distribution systems when actual utility data are proprietary. One workflow uses road network data, high-voltage substations, geo-referenced residences, and nearest-neighbor mappings from houses to road links and from candidate transformer sites to substations. For each road link, a secondary-network MILP selects transformer positions and builds a minimum-length radial forest connecting assigned residences to those transformers. For each substation, a primary-network MILP then selects a minimum-length tree on a road-graph proxy that connects the substation to its assigned transformers while respecting radiality, full coverage, and voltage-drop limits (Meyur et al., 2022).

Ensemble generation perturbs a near-optimal primary tree by repeatedly removing a random edge and reconnecting the resulting components via a different feasible path in the road graph, accepting only graphs that satisfy voltage, radiality, and coverage constraints. Validation against an actual Blacksburg network reports that more than 90% of nodes lie within $x_k/r_k=\alpha$9 of the real voltage, branch-power-flow histograms have $$S_k=S_{k+1}+s_{k+1}, \qquad V_k=V_{k+1}+\rho_k S_k,$$0, and degree, hop, and reach divergences are 0.02, 0.03, and 0.01, respectively. A related Virginia-scale construction reports 98% of nodes within $$S_k=S_{k+1}+s_{k+1}, \qquad V_k=V_{k+1}+\rho_k S_k,$$1 voltage deviation of the real grid, degree-distribution KL-divergence below 0.05, average path-length difference below 1%, voltage-profile RMSE below 0.005 pu, and line-loading histogram $$S_k=S_{k+1}+s_{k+1}, \qquad V_k=V_{k+1}+\rho_k S_k,$$2 $$S_k=S_{k+1}+s_{k+1}, \qquad V_k=V_{k+1}+\rho_k S_k,$$3-value above 0.9 (Meyur et al., 2022, Meyur et al., 2020).

4. Optimization, uncertainty, and distributed control

A large part of modern power-distribution research studies OPF under uncertainty and communication constraints. In the data-driven distributionally robust AC-OPF framework, forecast errors are modeled as zero-mean Gaussian variables with unknown variances lying in $$S_k=S_{k+1}+s_{k+1}, \qquad V_k=V_{k+1}+\rho_k S_k,$$4 confidence intervals estimated from historical data. Affine recourse allocates total active-power deviation across generators, and chance constraints on generation and voltages reduce to second-order conic forms by replacing each unknown variance with its worst-case upper endpoint. The resulting convex quadratically constrained program is reported to solve in under 1 s for 15–37 buses, about 1.1 s for 123 buses, and about 24 s for 8500 buses (Mieth et al., 2018).

Distributed methods target scalability and communication efficiency. In the unbalanced ADMM algorithm, the OPF is decomposed into local consensus subproblems so that each bus update reduces either to a closed-form solve or to projection of a $$S_k=S_{k+1}+s_{k+1}, \qquad V_k=V_{k+1}+\rho_k S_k,$$5 Hermitian matrix onto the PSD cone. Reported simulations on IEEE 13, 34, 37, and 123 feeders require approximately 300–600 iterations, with full-network solution in about 3 s when parallelized (Peng et al., 2015). Equivalent Network Approximation (ENApp) instead exchanges only upstream-equivalent voltages and downstream-equivalent complex powers. On IEEE 123-bus and IEEE 8500-node systems, ENApp converges in 4 and 17 macro-iterations, respectively, whereas an ADMM comparison required 272 and 756 macro-iterations (Sadnan et al., 2020).

Simulation-integrated distributed OPF uses a local linear model for optimization and a digital twin, such as GridLAB-D, for nonlinear projection of boundary voltages and flows before neighbor-to-neighbor exchange. On the IEEE 123-bus feeder partitioned into four areas, the reported performance is 4–7 macro-iterations, typical convergence in five communication rounds at 2 s per round, loss-minimization objective 44.959 kW versus 43.975 kW for nonlinear centralized OPF, DER maximization 5.163 MW versus 5.180 MW, and nodal voltages within $$S_k=S_{k+1}+s_{k+1}, \qquad V_k=V_{k+1}+\rho_k S_k,$$6 pu of the full-AC solution (Sadnan et al., 2022).

Voltage regulation has also been formulated as an AC-OPF rank-one SDP with a distributed dual-decomposition solver aligned to the electrical tree. The case studies on 5-, 34-, and 123-bus systems report approximately 20 iterations on the 5-bus example, average compute time of about 1.3 s per minute-update on the IEEE 34-bus feeder, and convergence within 300 iterations for about 80% of minutes on the IEEE 123-bus feeder, while maintaining voltages at all buses within $$S_k=S_{k+1}+s_{k+1}, \qquad V_k=V_{k+1}+\rho_k S_k,$$7 pu in the 34-bus study (Zhang et al., 2012).

A related consensus line treats power distribution as a supply-demand balancing problem over a connected undirected graph. There, each node computes desired net power, generation adjustments, and inter-node flows using only one-hop neighbor communication. The consensus filters recover a feasible capacity-weighted allocation satisfying $$S_k=S_{k+1}+s_{k+1}, \qquad V_k=V_{k+1}+\rho_k S_k,$$8, and the flow-control dynamics drive the residual net-power error to zero at every node (Kim et al., 2014).

5. Phase balance and optimal phase allocation

In unbalanced radial feeders, phase allocation is the problem of re-assigning the phase connections of loads and line segments in order to improve phase balance. A recent mixed-integer linear formulation defines continuous variables $$S_k=S_{k+1}+s_{k+1}, \qquad V_k=V_{k+1}+\rho_k S_k,$$9, $S_k=P_k+\alpha Q_k$0, and $S_k=P_k+\alpha Q_k$1, together with binary variables $S_k=P_k+\alpha Q_k$2 that indicate whether phase $S_k=P_k+\alpha Q_k$3 is present at bus $S_k=P_k+\alpha Q_k$4. The objective minimizes the sum of per-node voltage-unbalance terms and the total number of phases retained,

$S_k=P_k+\alpha Q_k$5

where $S_k=P_k+\alpha Q_k$6, subject to Lin3DistFlow voltage drops, nodal power balance, injection limits linked to $S_k=P_k+\alpha Q_k$7, total-injection conservation, phase-count limits, phase-consistency constraints, and voltage bounds (Gupta et al., 12 Jan 2025).

The principal unbalance metric is

$S_k=P_k+\alpha Q_k$8

A secondary metric identified in the same work is the Voltage Unbalance Factor,

$S_k=P_k+\alpha Q_k$9

Feeder	Case 2 result	Case 3 result
IEEE-13	$s_k=p_k+\alpha q_k$0 (67% reduction)	$s_k=p_k+\alpha q_k$1 (56% reduction)
IEEE-37	$s_k=p_k+\alpha q_k$2 (5% reduction)	$s_k=p_k+\alpha q_k$3 (9% reduction)
IEEE-123	$s_k=p_k+\alpha q_k$4 (34% reduction)	$s_k=p_k+\alpha q_k$5 (38% reduction)

These experiments use IEEE-13, IEEE-37, and IEEE-123 radial networks with openDSS data and three scenarios: per-phase capacity equal to spot load, $s_k=p_k+\alpha q_k$6 spot load, and $s_k=p_k+\alpha q_k$7 spot load. Reported solution times on a 16 GB MacBook M2 Pro are 0.1 s for IEEE-13, 3.7 s for IEEE-37, and 5.0 s for IEEE-123. The results also show that Case 3 is not universally best: on IEEE-13 it under-performs Case 2, illustrating the role of $s_k=p_k+\alpha q_k$8 in weighing phase-count penalties against unbalance reduction. The paper recommends offline deployment using meter/SCADA data, extension to dynamic or rolling-horizon settings, integration with GIS asset data, optional embedding of $s_k=p_k+\alpha q_k$9 decisions in remote switch controls, and post-solution validation against full-AC power flow (Gupta et al., 12 Jan 2025).

6. Probabilistic, computational, and instrumentation usages

Outside utility-grid studies, power distribution denotes several distinct distribution families. The Discrete Two-Sided Power distribution is introduced as a discrete analogue of the two sided power distribution. Its probability mass function and hazard rate function can assume bath tub, rectangular, trapezoidal, triangular, J, inverse J, U inverse U, strictly decreasing, and strictly increasing shapes, and its moment, reliability, and parameter-estimation properties are developed explicitly (Chakraborty et al., 2015).

In complex-valued signal modeling, the power-weighted noncentral complex Gaussian distribution defines

$\rho_k=r_k/V_0$0

with centroid $\rho_k=r_k/V_0$1, variance $\rho_k=r_k/V_0$2, and shape parameter $\rho_k=r_k/V_0$3. The induced amplitude and power laws provide a unified framework encompassing Rice, Nakagami, gamma, Rayleigh, and noncentral exponential special cases. The parameter $\rho_k=r_k/V_0$4 controls concentration near the origin versus arc-shaped phase diffusion, and maximum-likelihood fitting is performed numerically because no closed-form MLE exists for $\rho_k=r_k/V_0$5 in general (Nakashika, 27 Mar 2026).

In machine learning, the power distribution is the sequence-level target of power sampling:

$\rho_k=r_k/V_0$6

This distribution is the closed-form optimizer of KL-regularized reinforcement learning when the reward is the model’s own sequence-level log-probability, $\rho_k=r_k/V_0$7. It is also the target of power self-distillation, where teacher samples from $\rho_k=r_k/V_0$8 are amortized into supervised fine-tuning. Reported experiments on Qwen2.5-Math-7B over MATH500 give $\rho_k=r_k/V_0$9 for power sampling at $$v_{k,\phi}=v_{l,\phi}-2\sum_{\phi'}\left(r_{lk}^{\phi\phi'}p_{lk,\phi'}+x_{lk}^{\phi\phi'}q_{lk,\phi'}\right),$$0, compared with $$v_{k,\phi}=v_{l,\phi}-2\sum_{\phi'}\left(r_{lk}^{\phi\phi'}p_{lk,\phi'}+x_{lk}^{\phi\phi'}q_{lk,\phi'}\right),$$1 for standard decoding and $$v_{k,\phi}=v_{l,\phi}-2\sum_{\phi'}\left(r_{lk}^{\phi\phi'}p_{lk,\phi'}+x_{lk}^{\phi\phi'}q_{lk,\phi'}\right),$$2 for token-wise temperature; a distilled student with temperature decoding reaches $$v_{k,\phi}=v_{l,\phi}-2\sum_{\phi'}\left(r_{lk}^{\phi\phi'}p_{lk,\phi'}+x_{lk}^{\phi\phi'}q_{lk,\phi'}\right),$$3 (Tomihari et al., 6 May 2026).

The phrase also appears in detector engineering. The NOvA power distribution system powers avalanche photodiodes, thermoelectric coolers, front-end boards, and data concentrator modules for 344,064 readout channels in the far detector and 20,192 channels in the near detector. The architecture uses Wiener PL506 low-voltage supplies, Wiener MPOD HV-EX high-voltage supplies, relay racks, and compact power distribution boxes with single-point grounding at the PDB. Reported operation over four years exceeded 99.9% livetime, with no HV channel failures (Dukes et al., 2018).