Source-linked AI summary
Optimal Power Flow Using Graph Neural Networks
Damian Owerko, Fernando Gama, Alejandro Ribeiro
TL;DR
OPF requires cost-minimizing generator outputs under nonlinear AC power-flow constraints, while existing exact and approximate approaches face computational or scalability limits. The paper trains local, scalable graph neural networks through imitation learning to approximate IPOPT solutions from network states. Across IEEE-30 and IEEE-118 experiments, GNNs outperform comparable MLPs and deliver much faster, more scalable inference.
Problem
OPF is difficult because AC power-flow equations are non-convex, while exact IPOPT solutions are costly and do not scale to large networks.
Method
The paper trains a parametrized graph neural network with imitation learning to map measured network states to optimal generator outputs computed by IPOPT.
Results
GNNs outperform MLPs on IEEE-30 and IEEE-118, while inference takes tens of microseconds compared with seconds for IPOPT.
Takeaways & Limitations
Local GNN solutions that exploit grid structure provide an efficient approximation of the OPF solution across the two tested networks.
Abstract
from arXiv · showhide
Optimal power flow (OPF) is one of the most important optimization problems in the energy industry. In its simplest form, OPF attempts to find the optimal power that the generators within the grid have to produce to satisfy a given demand. Optimality is measured with respect to the cost that each generator incurs in producing this power. The OPF problem is non-convex due to the sinusoidal nature of electrical generation and thus is difficult to solve. Using small angle approximations leads to a convex problem known as DC OPF, but this approximation is no longer valid when power grids are heavily loaded. Many approximate solutions have been since put forward, but these do not scale to large power networks. In this paper, we propose using graph neural networks (which are localized, scalable parametrizations of network data) trained under the imitation learning framework to approximate a given optimal solution. While the optimal solution is costly, it is only required to be computed for network states in the training set. During test time, the GNN adequately learns how to compute the OPF solution. Numerical experiments are run on the IEEE-30 and IEEE-118 test cases.
1. INTRODUCTION
OPF is central to grid planning and operation but remains difficult because AC power-flow equations are nonlinear and non-convex. Existing approximations and learning methods have not adequately addressed scalability, motivating a local GNN imitation-learning approach.
- Motivation: OPF supports system planning, day-ahead pricing, and efficient allocation of generation capacity, but no fast and robust solution method is available.The paper notes that such a method could save tens of billions of dollars annually.
- Problem setting: OPF selects generator outputs to satisfy demand while minimizing production cost, but sinusoidal electrical behavior makes the problem difficult.Small-angle and constant-voltage assumptions yield a linearized alternative, but the passage introduces this only as an approximation.
- Computational challenge: AC OPF is non-convex because of nonlinearities in the power-flow equations and has been proven NP hard.Convex relaxations such as semidefinite programs are among the approaches discussed.
- Prior approaches: Prior machine-intelligence approaches include evolutionary, neural, fuzzy, swarm, and related methods, but none had been shown to work beyond the 30-node IEEE test case.The cited methods include evolutionary programming, genetic algorithms, artificial neural networks, simulated annealing, and particle swarm optimization.
- Prior approaches: A prior regression approach imitates ACOPF using information from selected, not necessarily neighboring, nodes, making it non-local.The paper contrasts this with architectures designed around network locality.
- Proposed direction: The paper uses imitation learning with a parametrized GNN to produce a local, scalable approximation of an optimal solution computed by a centralized interior-point solver.Experiments evaluate the framework on IEEE-30 and IEEE-118 power-system test cases.
2. OPTIMAL POWER FLOW
The paper models the grid as a weighted graph whose node measurements describe voltage and power states, then formulates OPF as constrained cost minimization. It targets a decentralized, scalable solution by learning a parametrized map from network states to optimal generator outputs.
- Grid representation: The power grid is represented as an undirected weighted graph whose nodes are buses and whose edge weights depend on line impedance.Weights use a Gaussian kernel and links below threshold ω are ignored.
- Grid representation: Each node state x_n contains voltage magnitude, voltage angle, active power, and reactive power, all of which can be measured locally.Across the network, these measurements are collected into vectors v, δ, p, and q.
- Generator representation: A selection matrix identifies the generator nodes among all buses and extracts their states from the full network state matrix.The generator-state collection is GX ∈ R^M×4.
- OPF formulation: OPF determines generator inputs that satisfy demand while minimizing generation costs subject to operational limits.Active and reactive powers depend nonlinearly on voltage magnitudes, voltage angles, and network topology.
- OPF formulation: DCOPF is convex under its approximation assumptions, whereas ACOPF is computationally costly because active and reactive power calculations are non-convex.The passage attributes the difference to the nonlinear power-flow relations.
- Learning objective: Imitation learning parameterizes a map Φ(X; H) and fits its parameters using a dataset of network states paired with optimal solutions.The desired objective is to learn the optimal solution in a decentralized and scalable manner.
- Learning objective: GNNs provide the chosen model because their locality and stability properties support scalable processing over graph-structured power-grid data.The paper specifically connects GNN stability properties with scalability.
3. GRAPH NEURAL NETWORKS
Graph neural networks process graph signals through localized shifts and learned graph convolutions, producing nonlinear representations whose parameter count is independent of network size. This locality and scalability motivate their use for imitating optimal OPF solutions.
- Graph shifts and convolutions: A graph shift updates each node from a linear combination of neighboring feature values, while repeated shifts collect information from progressively larger neighborhoods.The adjacency matrix W performs the shift, and W^kX can be computed through k repeated one-hop exchanges.
- Graph shifts and convolutions: Graph convolution combines shifted versions of an input signal using learned coefficient matrices H_k, preserving local computation.Multiplication by H_k changes feature combinations without changing the graph operation's locality.
- Graph neural networks: A GNN cascades graph-convolution layers with pointwise nonlinearities to map an input graph signal to an output representation.The layer states X_l use learned coefficients H_lk, with the final state X_L serving as the network output.
- Locality and scalability: Each GNN layer can be computed through repeated one-hop exchanges, and its filter parameters do not depend on the network size.The parameter count is |H| = Σ_l F_l−1F_lK_l, supporting a scalable architecture.
- Application to OPF: The paper trains this local, scalable GNN to imitate an optimal solution using a dataset of solver-generated examples.The target solution p* is obtained with an interior point solver, and the model parameters are learned over the training dataset.
4. NUMERICAL EXPERIMENTS
The experiments train and evaluate four architectures on synthetic IEEE-30 and IEEE-118 datasets, comparing graph-based and non-graph-based models against IPOPT-derived optimal outputs. GNNs outperform MLPs, while inference remains substantially faster than solving with IPOPT.
- Datasets: Datasets contain 8,016 IEEE-30 samples and 13,129 IEEE-118 samples, split 80% for training and 20% for testing.Samples use synthetic loads, sub-optimal grid states, and IPOPT-computed optimal generator outputs.
- Architectures and evaluation: The study compares global GNN, local GNN, global MLP, and local MLP architectures using MSE training and test-set relative error.The GNNs use graph structure, whereas the MLP baselines do not incorporate the underlying network structure.
- Results: 47% and 15% relative improvements occur on IEEE-30 for the global and local GNNs over their corresponding MLPs.The comparison is based on the reported relative performance improvements between matching GNN and MLP architectures.
- Results: 213% and 1082% relative improvements occur on IEEE-118 for the global and local GNNs over their corresponding MLPs.The larger-network results show especially large improvement for the local GNN relative to the local MLP.
- Computational cost: IPOPT requires 2.16s on IEEE-30 and 18s on IEEE-118, whereas GNN inference takes 48–56µs across the two datasets.The reported GNN timings are 48µs and 49µs on IEEE-30, and 50µs and 56µs on IEEE-118.
5. CONCLUSIONS
The paper approximates IPOPT-based OPF solutions with GNNs trained through imitation learning on IEEE-30 and IEEE-118 test cases. The experiments show that local solutions exploiting grid structure outperform comparable methods.
- Approach: GNNs approximate costly IPOPT OPF solutions from network states using the imitation learning framework.The approach uses local information processing while exploiting the grid’s network structure.
- Conclusion: Experiments on IEEE-30 and IEEE-118 show that local solutions exploiting the underlying grid structure outperform comparable methods.This conclusion is stated for the evaluated test cases and methods.
- Implication: The proposed architecture targets scalable and local OPF approximation rather than direct computation of the costly exact solution.The paper contrasts GNN approximation with the costly IPOPT solution.