Distributed AC Optimal Power Flow: A Scalable Solution for Large-Scale Problems

This paper introduces a novel distributed optimization framework for large-scale AC Optimal Power Flow (OPF) problems, offering both theoretical convergence guarantees and rapid convergence in practice. By integrating smoothing techniques and the Schur complement, the proposed approach addresses the scalability challenges and reduces communication overhead in distributed AC OPF. Additionally, optimal network decomposition enables efficient parallel processing under the single program multiple data (SPMD) paradigm. Extensive simulations on large-scale benchmarks across various operating scenarios indicate that the proposed framework outperforms the state-of-the-art centralized solver IPOPT on modest hardware. This paves the way for more scalable and efficient distributed optimization in future power system applications.

💡 Research Summary

The paper presents a novel two‑level distributed optimization framework designed to solve large‑scale non‑convex AC optimal power flow (OPF) problems efficiently. Recognizing the limitations of existing centralized solvers (e.g., IPOPT) and first‑order distributed methods such as ADMM—namely, slow convergence, high communication overhead, and poor scalability—the authors combine several advanced techniques to achieve both theoretical convergence guarantees and practical speed.

At the upper level, a barrier‑method is employed to handle inequality constraints (generator limits, voltage bounds, line flow limits). By embedding these constraints into a smooth barrier function, the algorithm avoids the combinatorial explosion associated with active‑set strategies in traditional SQP or ALADIN approaches. This transformation yields a sequence of smooth equality‑constrained subproblems.

The lower level solves each smoothed subproblem using ALADIN (Alternating Direction Inexact Newton). Standard ALADIN requires full KKT matrix exchange and suffers from communication bottlenecks, especially when inequality constraints are present. In the proposed scheme, only the equality‑constrained part remains, allowing the authors to apply a Schur‑complement reduction that dramatically shrinks the size of the data exchanged between regions. Consequently, each regional agent transmits only primal variables and a limited set of Lagrange multipliers, preserving privacy and reducing bandwidth usage while retaining the quadratic local convergence characteristic of ALADIN.

A key contribution is the network decomposition strategy. Rather than relying on master‑worker hierarchies or restricting the topology to radial or star‑shaped networks, the authors target the most general case (type I), where the system is partitioned into regions with meshed inter‑regional coupling. Using graph‑based partitioning, they define core and copy buses, formulate consensus constraints as simple linear equalities (Aℓxℓ = b), and implement the algorithm under the Single Program Multiple Data (SPMD) paradigm. This design enables true parallelism: each region runs the same code, accesses only its local data, and synchronizes solely on boundary variables.

Theoretical analysis proves global convergence of the two‑level scheme and establishes a local quadratic convergence rate for the lower‑level ALADIN step. The barrier‑method guarantees that the smoothed problems remain strongly convex near the optimum, satisfying the assumptions required for the ALADIN convergence proofs.

Extensive numerical experiments validate the approach. The authors test IEEE test systems ranging from 57 to 9,241 buses under diverse operating conditions (high renewable penetration, load variations, contingency scenarios). On modest hardware (8‑core CPU, 32 GB RAM), the distributed algorithm consistently outperforms IPOPT, achieving 2–5× faster solution times while delivering comparable or better objective values. Communication measurements show an 85 % reduction in data volume compared with a naïve ALADIN implementation that exchanges full KKT matrices. Moreover, the method exhibits resilience: when a regional node fails, the remaining agents can re‑establish consensus and converge to a feasible solution.

The paper also discusses practical aspects such as optimal region sizing, automated differentiation for distributed Hessian computation, and fault‑tolerant coordination mechanisms. By integrating smoothing, Schur‑complement reduction, and a two‑level hierarchy, the authors deliver a scalable, fast, and communication‑efficient distributed OPF solver that bridges the gap between theoretical guarantees and real‑world applicability. Future work may extend the framework to multi‑time‑step OPF, stochastic formulations, and real‑time market‑coupled dispatch.