# Optimal Power Flow Using Graph Neural Networks

**Authors:** Damian Owerko, Fernando Gama, Alejandro Ribeiro  
**arXiv:** [1910.09658](https://arxiv.org/abs/1910.09658)  
**Date:** 2019-10-21  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/1910.09658v1/

## Key points

- Optimal power flow (OPF) is a crucial optimization problem in the energy industry
- OPF aims to determine the most efficient power generation by grid generators to meet demand
- The optimality is evaluated based on the cost incurred by each generator
- Solving the OPF problem is challenging due to its non-convex nature caused by sinusoidal characteristics of electrical generation
- Small angle approximations have been used to simplify the problem, leading to a convex version known as DC OPF
- Approximate solutions proposed so far fail to scale effectively for large power networks
- Authors propose a novel approach using graph neural networks (GNNs) trained under imitation learning framework for approximating an optimal solution for OPF
- GNNs provide localized and scalable parametrizations of network data
- GNNs learn how to compute the OPF solution based on training experience during test time
- Numerical experiments conducted on IEEE-30 and IEEE-118 test cases validate the effectiveness of using GNNs in approximating optimal solutions for large-scale power networks.

## Summary

In the energy industry, optimal power flow (OPF) is a crucial optimization problem that aims to determine the most efficient power generation by grid generators to meet a given demand. The optimality is evaluated based on the cost incurred by each generator in producing this power. However, solving the OPF problem is challenging due to its non-convex nature caused by the sinusoidal characteristics of electrical generation. To simplify the problem, small angle approximations have been used, leading to a convex version known as DC OPF. However, this approximation becomes invalid when power grids are heavily loaded. Although several approximate solutions have been proposed, they fail to scale effectively for large power networks. In this paper titled "Optimal Power Flow Using Graph Neural Networks," authors Damian Owerko, Fernando Gama and Alejandro Ribeiro propose a novel approach using graph neural networks (GNNs) trained under the imitation learning framework to approximate an optimal solution for OPF. GNNs provide localized and scalable parametrizations of network data. The key idea behind their approach is that while computing the optimal solution for every network state can be computationally expensive during training time, GNNs learn how to compute the OPF solution based on its training experience during test time. To validate their method, numerical experiments are conducted on two standard test cases: IEEE-30 and IEEE-118. These experiments demonstrate the effectiveness of using GNNs in approximating optimal solutions for large-scale power networks. Overall, this paper introduces a promising approach that leverages graph neural networks and imitation learning to address the challenging task of solving OPF problems in energy grids efficiently and effectively.