MSAGNN

A Dual-Path GNN Integrating Static and Dynamic Graphs for Soft Sensor Modeling in Industrial Processes

Model and Results

The figure below illustrates the proposed Multi-Source Attention Graph Neural Network (MSAGNN) framework, which consists of industrial process data sampling, graph structure learning, attention mechanism fusion, GRU-based temporal modeling, and final output prediction.

Requirements

We recommend using Python 3.8 to 3.10, which has been tested and verified for this project.
Please ensure that the required dependencies are installed. You can install them all at once using the following command:

pip install -r requirements.txt

Dataset

This project uses the following industrial process datasets:

• TE Dataset (Tennessee Eastman Process): Commonly used for industrial process modeling and fault diagnosis research;
• DC Dataset (Distillation Column Process): A multivariate time-series dataset collected from a distillation system. All datasets are included in this repository and located in the ./data/ directory. No additional download is required.

Experiment Configuration and Execution

Parameter Configuration

The main experimental parameters are defined in the parser_args.py file and can be modified manually or overridden via command-line arguments.

Model Training

Run the training script using the command line:

python main.py --model_type AttGRU --dataset TE --target v10 --window_size 16 --horizon 1

Or use the provided shell script:

bash run.sh

Output

After training, the prediction results and logs will be automatically saved in the ./results/ directory.

Citation

@misc{wang2025MSAGNN,
title = {A Multi-Source Attention Graph Neural Network for Modeling Long and Short-Term Dependencies in Chemical Process Forecasting}, author = {Jian Long and bin wang and Haifei Peng and Hengmin Zhang}, year = {2025} }