[Paper Review] High-Fidelity Machine Learning Approximations of Large-Scale Optimal Power Flow

Minas Chatzos, Ferdinando Fioretto+2 more|arXiv (Cornell University)|Jun 29, 2020

Optimal Power Flow Distribution28 references42 citations

TL;DR

This paper presents OPF-DNN, a deep neural network approach integrated with Lagrangian duality to produce high-fidelity, millisecond-scale approximations of the AC-OPF for large-scale power systems (up to 3400 buses), with generator costs within 0.01% of optimum and strong constraint satisfaction.

ABSTRACT

The AC Optimal Power Flow (AC-OPF) is a key building block in many power system applications. It determines generator setpoints at minimal cost that meet the power demands while satisfying the underlying physical and operational constraints. It is non-convex and NP-hard, and computationally challenging for large-scale power systems. Motivated by the increased stochasticity in generation schedules and increasing penetration of renewable sources, this paper explores a deep learning approach to deliver highly efficient and accurate approximations to the AC-OPF. In particular, the paper proposes an integration of deep neural networks and Lagrangian duality to capture the physical and operational constraints. The resulting model, called OPF-DNN, is evaluated on real case studies from the French transmission system, with up to 3,400 buses and 4,500 lines. Computational results show that OPF-DNN produces highly accurate AC-OPF approximations whose costs are within 0.01% of optimality. OPF-DNN generates, in milliseconds, solutions that capture the problem constraints with high fidelity.

Motivation & Objective

• Motivate AC-OPF as a non-convex NP-hard problem and address increasing stochasticity in generation.
• Develop a learning-based surrogate (OPF-DNN) that produces AC-OPF setpoints while honoring physical constraints.
• Demonstrate high fidelity and scalability on large real-world networks (up to 3400 buses, 4500 lines).
• Integrate constraint handling via a violation-based Lagrangian relaxation within a DNN framework.

Proposed method

• Model the AC-OPF learning task as a constrained empirical risk minimization problem.
• Predict voltage magnitudes, angles, and generator active/reactive powers; infer line flows via Ohm's law to respect network physics.
• Use a violation-based Lagrangian relaxation to incorporate OPF constraints into the training loss, with a Lagrangian dual optimization to update multipliers.
• Employ a four-branch fully connected DNN architecture where each branch predicts v, theta, p^g, q^g; include loss components for constraints (L_c) and objective (L_o).
• Train with Adam on GPUs and update multipliers periodically to solve the Lagrangian dual (Algorithm 1).
• Evaluate on large French-system analogs to show millisecond prediction times and near-AC-OPF optimal costs.

Experimental results

Research questions

• RQ1 Can a DNN-based model approximate AC-OPF solutions for large-scale power systems with high fidelity while respecting physical constraints?
• RQ2 How well do violation-based Lagrangian relaxations enable learning models to satisfy AC-OPF constraints during training and prediction?
• RQ3 What is the trade-off between prediction speed and solution optimality/feasibility when scaling to networks with thousands of buses?
• RQ4 How does OPF-DNN perform on real-world large networks (up to ~3400 buses) in terms of accuracy, feasibility, and computational resources?

Key findings

• OPF-DNN achieves millisecond prediction times for large-scale AC-OPF instances.
• The approach yields AC-OPF objective values very close to the true optimum, with cost differences in the reported experiments being minimal relative to AC-OPF baselines.
• Constraint satisfaction is highly feasible with the proposed method, with most constraints satisfied at near 100% and small mean violations for the rest.
• Indirect flow predictions via predicted voltages and Ohm’s law improve accuracy for line flows and help satisfy AC-OPF constraints.
• Compared to unconstrained baselines, the violation-aware training (OPF-DNN) converges more reliably and yields better feasibility and objective performance on large networks.
• Training on French-system-scale networks (up to 3,411 buses, 4,499 lines) demonstrates scalability and practical applicability.