Refining Graphical Neural Network Predictions Using Flow Matching for Optimal Power Flow with Constraint-Satisfaction Guarantee

The DC Optimal Power Flow (DC-OPF) problem is fundamental to power system operations, requiring rapid solutions for real-time grid management. While traditional optimization solvers provide optimal solutions, their computational cost becomes prohibitive for large-scale systems requiring frequent recalculations. Machine learning approaches offer promise for acceleration but often struggle with constraint satisfaction and cost optimality. We present a novel two-stage learning framework that combines physics-informed Graph Neural Networks (GNNs) with Continuous Flow Matching (CFM) for solving DC-OPF problems. Our approach embeds fundamental physical principles–including economic dispatch optimality conditions, Kirchhoff’s laws, and Karush-Kuhn-Tucker (KKT) complementarity conditions–directly into the training objectives. The first stage trains a GNN to produce feasible initial solutions by learning from physics-informed losses that encode power system constraints. The second stage employs CFM, a simulation-free continuous normalizing flow technique, to refine these solutions toward optimality through learned vector field regression. Evaluated on the IEEE 30-bus system across five load scenarios ranging from 70% to 130% nominal load, our method achieves near-optimal solutions with cost gaps below 0.1% for nominal loads and below 3% for extreme conditions, while maintaining 100% feasibility. Our framework bridges the gap between fast but approximate neural network predictions and optimal but slow numerical solvers, offering a practical solution for modern power systems with high renewable penetration requiring frequent dispatch updates.

💡 Research Summary

**
The paper tackles the computational bottleneck of solving the DC Optimal Power Flow (DC‑OPF) problem in real‑time power‑system operation. Traditional interior‑point or sequential quadratic programming solvers guarantee optimality but become too slow for large networks that must be re‑solved hundreds of times per day as renewable generation and load fluctuate. Recent machine‑learning approaches promise rapid inference but typically suffer from constraint violations, large optimality gaps, and poor generalization to stressed operating conditions.

To bridge this gap, the authors propose a two‑stage learning framework that first generates a feasible dispatch with a physics‑informed Graph Neural Network (GNN) and then refines that dispatch toward optimality using Continuous Flow Matching (CFM), a simulation‑free continuous normalizing flow technique.

Stage 1 – Physics‑Informed GNN
The power grid is modeled as an undirected graph whose nodes represent buses and edges represent transmission lines. Node features consist of the real‑power demand at each bus; optional edge features encode line parameters. A two‑layer Graph Convolutional Network (GCN) with 128‑dimensional hidden states processes the graph, after which a two‑layer multilayer perceptron predicts a raw generation vector for all generators. Because raw neural‑network outputs are unconstrained, the authors introduce two projection mechanisms. During training a differentiable “soft‑clamp” with a temperature parameter (τ = 0.05) allows gradients to flow while keeping outputs within generator limits. Power‑balance is enforced by proportionally adjusting each generator’s output according to its capacity. At inference time a hard iterative projection (Algorithm 1) guarantees exact satisfaction of generator limits and the global power‑balance equation within a tolerance of 0.005 MW.

The loss function combines six physics‑aware terms:

1. Cost‑gap loss – relative difference between predicted cost and the optimal cost obtained from a conventional solver.
2. Economic‑dispatch loss – variance of marginal costs across generators; at optimality all marginal costs equal the system marginal price λ, so the variance should be zero.
3. KKT complementarity loss – penalizes violations of the complementary slackness conditions at lower and upper generation bounds using ReLU‑based terms.
4. Power‑balance loss – L2 norm of the mismatch between total generation and total demand.
5. Limit‑violation loss – ReLU penalties for any generator output outside its prescribed bounds.
6. Direct‑cost loss – the absolute generation cost of the predicted dispatch, encouraging the network to learn low‑cost solutions directly.

A curriculum learning schedule gradually shifts emphasis from feasibility (high weights on balance and limits) in the first third of training epochs, to physics‑based optimality (increased economic and KKT weights) in the middle third, and finally to pure cost minimization (large cost‑gap and direct‑cost weights) in the last third. This staged weighting stabilizes training and guides the network toward solutions that are both feasible and near‑optimal.

Stage 2 – Continuous Flow Matching Refinement
Given a feasible initial dispatch (p^{(0)}_g) from Stage 1 and the true optimal dispatch (p^{}_g) from the solver, the authors define a linear interpolation path in generation space:
(p_g(t) = (1-t)p^{(0)}_g + t p^{}_g,; t\in