Stochastic Distribution Network Reconfiguration under Load Uncertainty

1 Introduction

• Reconfiguration changes switch statuses to reduce technical losses while maintaining load supply and network radiality.Source: §1 · p. 1
• A deterministic configuration may be overly tailored to one known demand level, whereas stochastic reconfiguration evaluates a common topology under multiple loading scenarios.Demand variability is linked to distributed micro- and minigeneration and electric vehicles.Source: §1 · p. 1
• The study makes topology decisions in the first stage and determines power flows and other operational variables separately for each scenario in the second.Source: §1 · p. 2
• The deterministic equivalent is a multiscenario MISOCP incorporating LinDistFlow, a convex quadratic loss relation, radiality constraints, and an optional switching penalty.The model is solved directly with Gurobi.Source: §1 · p. 2

2 Literature Review

• The reviewed studies progress from deterministic loss-reducing topology selection toward formulations using probabilities, operational data, and multiple uncertainty sources.Source: §2 · p. 3
• The reviewed literature lacks a uniform protocol for comparing deterministic and stochastic solutions while holding the network representation, objective, and modeling assumptions constant.Source: §2 · p. 3
• The study initially restricts uncertainty to load and keeps the topology common across scenarios, with operational variables adjusted after demand is realized.Source: §2 · p. 3
• Its controlled scope enables comparisons of expected losses, worst-case scenario losses, and electrical violations within the same mathematical framework.The stated contribution is a reproducible analysis of explicitly accounting for demand uncertainty, rather than adding more uncertain sources.Source: §2 · p. 3

3 Fundamental Concepts

• The approach is a quantitative modeling and computational study that adds demand scenarios to a validated deterministic representation while retaining the candidate network and primary electrical criterion.Source: §3 · p. 4

3.1 Graph-theoretic representation of the distribution network

• Radiality is enforced through one incoming branch per nonroot bus, |N| −1 closed branches, and dynamic elimination of cycles.The resulting topology is checked again during electrical evaluation.Source: §3.1 · p. 4

3.2 Concepts related to deterministic reconfiguration

• Deterministic reconfiguration selects a radial topology for known demand using a convex power-flow approximation, then evaluates the fixed topology with approximate radial power flow.The evaluation reports approximate losses, voltage profiles, and violations, not exact AC power flow results.Source: §3.2 · pp. 4–5

3.3 Concepts related to stochastic reconfiguration

• In stochastic reconfiguration, one topology is chosen before demand is known, while flows, voltages, and auxiliary loss variables are scenario-specific recourse decisions.The topology variable has no scenario index, while operational variables are replicated across scenarios.Source: §3.3 · p. 5

3.4 Illustrative example

• The illustrative example contrasts choosing a topology for one known loading condition with choosing a common topology evaluated across low, nominal, and high-load scenarios.The example assigns probabilities 0.25, 0.50, and 0.25 to the three scenarios.Source: §3.4 · p. 6

3.5 Description of the problem addressed

• The problem selects radial branch statuses before demand is realized, evaluates scenario-specific electrical operation, and primarily minimizes expected approximate losses.An optional penalty can represent changes relative to the initial configuration.Source: §3.5 · p. 6

3.5.1 Loss calculation and approximation

• Topology optimization uses an approximate conic loss representation, while a separate approximate radial power-flow evaluation computes losses after the topology is fixed.Neither stage substitutes for high-fidelity nonlinear AC validation.Source: §3.5.1 · pp. 6–7

4 Mathematical Model

• The deterministic equivalent is classified as a MISOCP, with binary topology variables and convex quadratic constraints; operational variables are scenario-specific.In the Gurobi interface, the relations are entered as convex quadratic constraints in a MIQCP.Source: §4 · p. 7

4.1 Optimization problem formulation

• The model uses a rooted directed candidate graph, with buses, branches, switchable branches, and identified cycles defining its topology.Branch orientation structures the tree and balance equations; flows may be positive or negative.Source: §4.1 · p. 7

4.1.1 Objective function

• The objective minimizes expected electrical performance across scenarios and can add a weighted switching penalty, but γ = 0 in the reported experiments.Thus, the switching penalty does not affect the reported results.Source: §4.1.1 · p. 8

4.1.2 Topological constraints and radiality

• Unique-predecessor, topology-cardinality, and cycle inequalities jointly enforce the rooted radial structure.The cycle inequalities prevent all branches of any identified cycle from remaining closed.Source: §4.1.2 · p. 8

4.1.3 Power balance by scenario

• For each scenario, nonroot-bus balances omit explicit loss terms, while the root’s injection is determined implicitly by flows leaving it.Source: §4.1.3 · p. 8

4.1.4 Coupling topology, flows, and losses

• Open branches have zero flows and auxiliary loss variables; a convex second-order-cone relation approximates flow magnitude and losses.The nonnegative loss coefficient in the objective encourages the auxiliary variable to take its smallest feasible value.Source: §4.1.4 · p. 9

4.1.5 Voltage drops and bounds

• The model fixes root voltage and uses LinDistFlow voltage drops on closed branches, relaxing the voltage-drop relation when a branch is open.Source: §4.1.5 · p. 9

4.1.6 Switching penalty

• An optional weighted penalty can count or price switching changes, but with γ = 0 the experiments compute switching operations only after optimization.The number of changes is obtained by comparing the selected branch statuses with their initial statuses.Source: §4.1.1 · p. 8, §4.1.6 · p. 9

4.1.7 Compact form

• With finite scenarios, binary branch statuses are the only discrete variables and convexity is confined to the second-order-cone inequalities, yielding a multiscenario MISOCP.The deterministic case is recovered with one scenario.Source: §4.1.7 · pp. 9–10

• First-stage topology and optional switching-penalty decisions are common to all scenarios, with radiality enforced by predecessor, cardinality, and cycle constraints.Source: §4.2 · p. 10
• Second-stage electrical variables vary by scenario, while the shared topology prevents selecting a different configuration after demand is observed.This implements nonanticipativity.Source: §4.2 · p. 10
• The objective weights electrical performance by scenario probabilities and permits a switching cost, but switching does not influence the reported decisions because γ = 0.The selected topology is evaluated afterward using an approximate Baran–Wu radial method.Source: §4.2 · p. 10

• The model is single-period, selects one topology shared across scenarios, treats only demand as uncertain, and assumes scenario probabilities are known.Source: §4.3 · p. 10
• Power balances omit branch losses, the current relation omits the sending-end voltage factor, and LinDistFlow drops the quadratic loss term in voltage drops.Source: §4.3 · p. 10
• Activation bounds do not provide detailed equipment thermal limits, and ex post electrical evaluation uses approximate radial power flow rather than nonlinear AC power flow.Source: §4.3 · p. 10
• The objective is risk-neutral, and the supported switching penalty is disabled in the experiments; proposed extensions include heterogeneous scenarios, multiperiod models, risk measures, and fuller validation.The authors also propose distributed generation, storage, systematic γ analysis, and decomposition for larger scenario sets.Source: §4.3 · p. 10

5 Solution Techniques

• A monolithic Gurobi solution of the deterministic equivalent uses shared topology variables, scenario-specific continuous variables, and lazy cycle constraints.Source: §5 · p. 10

5.1 Techniques used in the literature

• The literature uses decomposition for two-stage robust models, monolithic MISOCP, metaheuristics, and sample average approximation across differing uncertainty and network settings.The paper chooses a moderate scenario set and exact commercial solver to record bounds, gap, status, and time.Source: §5.1 · p. 11

5.2 Model construction and solution with Gurobi

• The implementation constructs and solves the deterministic equivalent in Python through gurobipy, recording solver status, objective, bound, MIP gap, and computation time.Source: §5.2 · p. 11

5.3 Warm start from the initial configuration

• The full warm start initializes topology and, when the initial network is a rooted directed tree, scenario flows, loss auxiliaries, and voltages from that configuration.It changes neither the feasible set nor the optimality criterion; it may accelerate discovery of incumbents.Source: §5.3 · p. 11

5.4 Dynamic separation of cycle elimination constraints

• Cycle inequalities are generated on demand for cycles detected in candidate integer solutions, avoiding advance enumeration of all simple cycles.Cycle detection uses an undirected support graph, with native or NetworkX backends.Source: §5.4 · pp. 11–12

5.5 Solution workflow and ex post evaluation

• After optimization, the fixed topology is evaluated separately in each scenario with approximate Baran–Wu radial power flow to report losses, voltages, and violations.The evaluator checks radiality and root consistency; switching counts are postprocessed and do not affect the objective.Source: §5.5 · p. 12

5.6 Computational challenges

• Scenario growth replicates operational variables and constraints, while oversized coupling bounds weaken relaxations and overly restrictive bounds may exclude feasible solutions.Dynamic cycle separation reduces initial constraints but adds callback overhead; the study records solution quality and uses separate approximate evaluation.Source: §5.6 · p. 12

6 Computational Experiments

• Reported losses are from ex-post approximate Baran–Wu radial power flow, not directly from the MISOCP objective.Current-capacity data were not uniformly available, so overload metrics were omitted from the main comparison.Source: §6 · p. 13, §6.1 · p. 13

6.1 Computational environment and implementation

• The implementation solves a monolithic MISOCP, with switching penalties disabled and cycle elimination handled through lazy constraints.The experiments ran on an Apple M2 computer with 16 GB of memory; optimization voltage bounds were 0.60–1.05 pu, while ex-post violations use stricter 0.95–1.05 pu bounds.Source: §6.1 · p. 13

6.2 Test instances

• The 12 test networks span 17 to 10,561 nodes and 16 to 10,728 physical branches.The set ranges from small instances to networks with thousands of buses and hundreds of effective topology decisions.Source: §6.2 · p. 13

6.3 Demand scenarios

• Three common scenarios globally scale active and reactive demand while preserving relative load locations and each bus’s power factor.The probability-weighted mean load is 103% of nominal; the nominal deterministic case instead uses 100%.Source: §6.3 · pp. 13–14

6.4 Experimental protocol and metrics

• Each network’s initial topology and the optimized common topology were evaluated across all scenarios, recording expected and maximum losses, voltage outcomes, switching operations, and solver indicators.The stochastic MISOCP uses one first-stage topology with scenario-specific operational variables; EEV evaluates the mean-load deterministic topology in the original scenarios.Source: §6.4 · p. 14

6.5 Overall loss results

• 3.23%–65.23%: expected losses fell in all 12 instances, with a mean reduction of approximately 31.76%.Absolute losses in the nominal deterministic and stochastic expected-loss columns are not directly comparable; compare percentage reductions under each criterion.Source: §6.5 · p. 14, Table 2 · p. 15
• Deterministic and stochastic percentage reductions coincided to reported precision in almost all networks; bus_1579_13 was the visible exception after reaching the time limit.The paper relates this pattern to the global load scaling used to construct the scenarios.Source: §6.5 · p. 14

6.6 Electrical and operational performance

• 1,589 to 714: voltage violations fell by approximately 55.1% across the 12 networks and three scenarios, and maximum loss decreased in every instance.Five networks had zero violations in the ex-post evaluation.Source: §6.6 · p. 15, Table 3 · p. 15
• With the switching penalty disabled, switching operations were not minimized and reached 472 in bus_3115_24.The paper identifies the optional switching penalty as useful for future experiments.Source: §6.6 · p. 15

6.7 Value of the stochastic solution

• In seven instances the stochastic and expected-value topologies matched with VSS = 0; among ten instances where both were optimal, absolute relative VSS was at most 0.0017%.Small negative reported VSS values are treated as numerical equivalence because the ex-post loss evaluation differs from the approximate conic objective and uses numerical tolerances.Source: §6.7 · pp. 15–16, Table 4 · p. 16

6.8 Comparison of deterministic and stochastic topologies

• Across the evaluated instances, the additional benefit of stochastic topologies over a well-chosen deterministic topology was small; in bus_3115_24, expected loss fell from 2143.146 to 2138.837 kW, an additional reduction of approximately 0.201%.Source: §6.8 · p. 16

6.9 Scenario-specific behavior

• High_load had the highest losses and lowest voltage in every network, consistent with scenarios differing only by a global load multiplier.Voltage violations fell from 13 to 0 in low_load, 40 to 19 in nominal_load, and 1536 to 695 in high_load after reconfiguration.Source: §6.9 · p. 16
• The scenario factor 1.105 is topology-independent, so nominal-loss and expected-loss minimization can yield the same topology ordering when voltage constraints do not alter competitive choices.This scaling structure explains the coincident percentage reductions and nearly zero VSS.Source: §6.9 · pp. 16–17

6.10 Computational performance and scalability

• 11 of 12 instances reached OPTIMAL status; bus_1579_13 stopped at 600 s with an approximately 2.43% gap.Among the 11 certified instances, median solution time was approximately 1.45 s.Source: §6.10 · p. 17
• Runtime did not increase monotonically with network size; bus_2011_15 required about 318 s, while the 10,561-node bus_10476_84 solved in about 34 s.Effort also depended on candidate-branch structure, relaxation quality, topology alternatives, and interactions among warm starts, cycle cuts, and integer search.Source: §6.10 · p. 17
• bus_1579_13 had an approximately 60.71% expected-loss reduction and no voltage violations in subsequent evaluation, but global optimality was not certified within the limit.The authors identify stronger relaxations and decomposition as future directions for this case.Source: §6.10 · p. 17

6.11 Discussion of results

• The experiments distinguish substantial benefits from reconfiguration itself from the small incremental value of stochastic modeling under scenarios that preserve spatial load distribution.Seven networks had identical RP–EV topologies; differences in the others were numerically small.Source: §6.11 · p. 17
• Scenarios that alter spatial demand distribution, regional correlations, or active and reactive power variations are needed to test conditions where stochastic information may change the first-stage decision.The current experiments validate the formulation across varied network sizes while identifying the limitation of global-scaling scenarios.Source: §6.11 · p. 17

7 Concluding Remarks

• The two-stage model shares topology decisions across scenarios and determines flows, voltages, and auxiliary variables separately, using a multiscenario MISOCP with radiality and LinDistFlow.The model was solved monolithically with Gurobi using a warm start and dynamic cycle separation through lazy constraints.Source: §7 · p. 17
• Expected losses fell in all 12 networks, with reductions from 3.23% to 65.23% and an average reduction of approximately 31.76%.The cumulative voltage violations in ex post evaluation decreased from 1589 to 714, approximately 55.1%.Source: §7 · pp. 17–18
• Seven networks had identical RP and EV topologies with VSS = 0, while differences in the others were very small under global load scaling.The scenarios used factors 0.70, 1.00, and 1.30 while preserving relative spatial distributions of active and reactive demand.Source: §7 · p. 18
• The formulation supports switching penalties, but the reported experiments set γ = 0; topology selection can still reduce losses and improve voltage profiles as demand varies.The study did not quantify the economic or social effects of those electrical improvements.Source: §7 · p. 18
• Uncertainty was limited to three global-scaling demand scenarios, with no out-of-sample evaluation, and the electrical model and validation rely on DistFlow/LinDistFlow approximations.The optimization voltage bounds were more permissive than the subsequent evaluation range, and branch current capacities were not uniformly available.Source: §7 · p. 18