GNN-PowerFlow: Innovative Application of Graph Neural Networks in Power System Power Flow Calculation

This project is a master's thesis from Eindhoven University of Technology, exploring the application of Graph Neural Networks (GNN) in AC power system power flow calculation. The core is to improve prediction accuracy by integrating grid topology and conduct comparative analysis with traditional Multi-Layer Perceptrons (MLP). The project code is open-sourced on GitHub (link: https://github.com/mukhlishga/gnn-powerflow) and implemented using the PyTorch Geometric framework.

Power system power flow calculation is a core task. Traditional methods rely on numerical iteration, and the cost increases with the scale of the grid. The integration of renewable energy increases complexity, requiring fast and accurate methods. Traditional neural networks (such as MLP) ignore data structure relationships, while the power grid is naturally a graph structure (buses as nodes, lines as edges), making it suitable for GNN applications. Core problem: Under the same model complexity, can incorporating grid topology into machine learning models improve the prediction accuracy of power flow calculation?

In the power grid, nodes are buses (features include voltage magnitude, phase angle, active/reactive power), and edges are transmission lines (features such as current, resistance, reactance). Buses are divided into three types: Slack buses (known V/δ, unknown P/Q), PV buses (known P/V, unknown Q/δ), and PQ buses (known P/Q, unknown V/δ). Traditional methods solve unknowns using equations, while GNN predicts unknowns through known variables.

The core of GNN is message passing: nodes aggregate neighbor information. The adjacency matrix A encodes connectivity to enable message passing (e.g., summation). GCN uses average aggregation (preprocessed via degree matrix). Multi-layer GNN can achieve global information propagation; for example, a 9-node graph extracts features through two layers of transformation.

The experiment uses the PyTorch Geometric framework to compare multiple GNN architectures with MLP of the same complexity. The results show that GNN improves accuracy by leveraging topology, while MLP performs poorly due to lack of structural awareness, verifying the core hypothesis.

Academic contribution: Provides empirical evidence for the application of GNN in power system power flow calculation, and systematically compares GNN with traditional neural networks. Practical value: Supports real-time power flow prediction, large-scale grid analysis, renewable energy integration, and grid optimization planning.

The project code is based on PyTorch Geometric, including experimental notebooks and data processing workflows. The open-source nature promotes knowledge sharing and reproducibility, contributing to the development of the machine learning field in power systems.