Overview / Usage

Each year billions of dollars are lost due to inefficient operation of the electrical grid. Optimal Power Flow is the problem of finding an output of each generator in an electrical grid that minimizes production cost. It is NP-hard due to it's nonlinear constraints that arise due to the trigonometric equations involved in computing AC power flow. The problem is only becoming more complex as renewable energy and other distributed energy resources are incorporated into the grid.

Most commonly, interior point methods are used to solve the problem. However, they suffer from scalability issues for large electrical grids. Thus, there has been a growing interest in learning approaches that can, at the cost of long training times, dramatically increase computational efficiency once deployed. However, deep neural networks suffer from instability issues and do not scale well to large networks. In the past we shown that graph neural networks (GNNs) can outperform such DNNs in a supervised setting. Currently we are working on a novel unsupervised approach that we hope will train GNNs that could viably be used as ACOPF solutions.

Methodology / Approach

We propose a novel unsupervised optimizer for constrained problems inspired by log-barrier methods. There are two parts to our gradient updates. First, we use the optimal power flow constraints, to take a gradient descent step towards the feasible space. Second, we perform gradient descent on the convex cost function, associated with optimal power flow.

All the code is written in PyTorch and NumPy and is publically available on the github repository.

Technologies Used

We are using Intel DevCloud for data processing and model training.

Repository

https://github.com/Damowerko/OPF

Other links

• ICASSP2020 Presentation
• Optimal Power Flow Using Graph Neural Networks