MFGAT

A Knowledge-Guided Graph Reasoning Framework for Physics-Consistent Multi-physics Field Prediction in Converter Transformers Table of Contents File Structure Data Preprocessing Flow Dataset Splitting Model Architecture Core Script Description Usage Installation Environment Note on Dataset Size and Availability This research presents a novel framework named the Multi-Feature Graph Attention Network (MFGAT), designed for the efficient and accurate prediction of the thermo-fluid field within converter transformers. The framework integrates physical knowledge from traditional Computational Fluid Dynamics (CFD) methods with Graph Neural Networks (GNN) , significantly enhancing both computational efficiency and prediction accuracy. This repository contains the code implementation of the MFGAT model proposed in the paper, which handles ultra-large-scale heterogeneous graph data and provides two schemes for evaluating generalization capabilities: Multi-Condition Generalization and Regional Generalization.。 1. File Structure The core directory structure of this project is as follows: main.py #Training script predict.py #Prediction and evaluation script checkpoints #Trained model checkpoints dataset(operating conditions split) #Dataset Directory (Multi-Condition Split) raw #Original physical field data files (processed by data_handle) labels #Label files subgraphs #Physical subgraph files test #Test set symbolic links train #Training set symbolic links val #Validation set symbolic links dataset(regional split) #Dataset Directory (Regional Split) test #Test set subgraph files and Labels train #Train set subgraph files and Labels val #Validate set subgraph files and Labels data_handle #MATLAB scripts for data preprocessing logs #Log files models #Python scripts related to the model data.py #Data loading and normalization model.py #MFGAT model definition