Pioneering Generalizable Graph Neural Network Solutions for Optimal Power Flow in Modern Power Grids.

Introduction

Power grids are among the largest, most complex, and most critical infrastructures in modern society. Solving alternating current optimal power flow (AC-OPF) and its variants is essential for ensuring economically efficient and reliable power markets. This process involves determining optimal unit commitment, power dispatch, and related quantities while subjecting to physical constraints.

However, the non-convex and non-linear characteristics of AC-OPF pose significant challenges in finding optimal solutions with high efficiency. In recent years, considerable research has been conducted on solving AC-OPF using mathematical and metaheuristic approaches. Nevertheless, many of these methods are challenging to expedite due to the need for repetitive solving when conditions change, and they often lack scalability.

Machine learning-based approaches have been widely adopted to address AC-OPF due to their potential to accelerate the solving process and their adaptability to large-scale AC-OPF problems. Deep neural network models have been utilized to directly predict solutions or reduce inactive constraints. Graph neural network models are alternatively incorporated to take grid topology into consideration, thereby making the well-trained model adaptive to rapid topology changes.

The feasibility of solutions is further considered to avoid constraint violations; a typical method is to incorporate penalties for constraint violations into the loss function. Aims and Objectives In summary, this research aims to address the following gaps in current AC-OPF research:

- Adaptability to topology change. Current methodologies often struggle to dynamically adjust to changes in grid topology due to planned addition or removal of grid components, network topology switching, or unplanned outages. This research will explore advanced techniques to enhance the adaptability of machine learning-based AC-OPF methods.

- Feasibility of solutions. Machine learning-based methods achieve speed improvements without the need to repeatedly solve similar AC-OPF problems, but they cannot guarantee the feasibility of the solutions, i.e., they may not always satisfy the necessary constraints. This research will investigate methods to guarantee the feasibility of generated solutions.

- Scalability to large-scale grids. Many of current research on AC-OPF focus on small-scale grids. As power systems grow in complexity and size, scalability becomes increasingly important. This research will focus on developing scalable algorithms that can efficiently handle the computational demands of large-scale AC-OPF problems.

- Computation efficiency of solution generation. The computation efficiency of addressing the AC-OPF problem is vital for real-time applications in power systems. This project will aim to improve the computational efficiency of AC-OPF algorithms through the integration of hybrid techniques that combine machine learning with traditional optimization methods.

Applications and Impacts

This research will have a significant impact on electricity market operations and power system management. Enhancing the adaptability, feasibility, scalability, and computational efficiency of solving AC-OPF, will facilitate real-time operations, particularly in the context of integrating renewable energy sources and accommodating flexible demand. Consequently, this research is expected to yield benefits such as increased grid reliability, optimized operational costs, reduced carbon emissions, and a more resilient and economically efficient power infrastructure.

This project falls within the EPSRC research areas of Energy and Decarbonisation, as well as Engineering and Mathematical Sciences.