arXiv:2609.40264
Two-Stage Quantum-Classical Distribution Network Reconfiguration via Cycle-Edge Encoding
Nowar Alashkar, Cade Kennedy, Phillip C. Lotshaw, Shaked Regev, Miguel Angel Lopez-Ruiz, Ananth Kaushik, Paul Smith, Claudio Girotto, Martin Roetteler
Abstract
Distribution network reconfiguration (DNR) is a combinatorial optimization problem that seeks a low-loss network topology subject to topological and electrical operating constraints. In this work, we propose a two-stage quantum-classical method for DNR. First, an iterative linear-ramp Quantum Alternating Operator Ansatz (LR-QAOA) searches topologically feasible subspaces constructed using a novel cycle-edge encoding. The encoding guarantees that every represented configuration is a spanning tree while allowing the size and coverage of each subspace to be adjusted to the available qubit budget. Second, the resulting sample distribution is used to guide a mixed-integer second-order-cone programming (MISOCP) solver. The method is evaluated on six distribution networks ranging from 33 to 417 buses through ideal simulation and execution on trapped-ion quantum hardware. For the 417-bus network, quantum hardware sampling identifies a configuration within approximately $1.4\%$ of the literature reference loss. Across all six test systems, the MISOCP solver guided by hardware-derived distributions reaches an incumbent within $1\%$ of the target with a lower median solver time than the corresponding unguided solver. These findings establish a concrete framework for applying current quantum hardware to DNR, with the potential to become increasingly competitive as hardware improves.
AI Summary
The paper studies whether quantum-generated samples can guide classical optimization for distribution network reconfiguration: minimizing electrical losses subject to radiality, connectivity, and operating constraints. It proposes a two-stage workflow combining iterative linear-ramp QAOA with a mixed-integer second-order-cone programming (MISOCP) solver. A cycle-edge encoding partitions cycle edges into disjoint blocks, with exactly one edge opened per block. The paper proves that every permitted assignment yields a spanning tree and that a finite collection of partitions covers all spanning trees; this does not establish efficient coverage. Each iteration constructs a qubit-budgeted subspace around a reference tree, fits a quadratic loss surrogate to classical power-flow evaluations, and samples constraint-preserving QAOA circuits. Selected samples update the reference tree. The final samples provide a MIP start and variable hints to the classical solver. Tests use six networks spanning 33–417 buses, with 29 qubits, two QAOA layers, and 1,000 shots per circuit, using ideal simulation and trapped-ion hardware. For the 417-bus network, the best simulated and hardware-sampled losses are 586.66 and 589.56 kW, respectively, versus a literature reference of 581.57 kW. Hardware-guided MISOCP reduces median solver time to a power-flow-validated solution within $1\%$ of the reference across all six networks; for 417 buses, it decreases from 1,415.47 to 34.71 seconds across 20 solver seeds. Quantum-stage overhead prevents an end-to-end runtime advantage. The paper notes heuristic, nonmonotonic search, noise-induced infeasible samples, restricted subspace coverage, and uncertified reference optima. Larger-instance crossover remains untested.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 10 pages, 8 figures