Application of graph neural networks in modeling volatile matter generation and internal gas pressure behavior during coking of coking coal blends

• A novel CFD-graph neural network hybrid approach for predicting the rate of volatile matter generation during coal coking. • A deep neural network-based feature augmentation method to improve temperature-dependent feature representation. • Graph data was constructed from coking process characteristics for the prediction of the volatile matter generation rate. • The internal gas pressure was estimated based on the deep neural network and graph attention network models. As a traditional carbon-intensive and high-energy consumption industry, the coking industry can significantly benefit from digital transformation to optimize coke production. Volatile matter (VM) generation is a critical stage in the coking process, and its efficient and accurate prediction plays a crucial role in enhancing coke quality, recycling coke oven gas, and ensuring production safety. However, it is difficult to estimate the VM generation during coking due to the nonlinearity of the process and the difficulty of measurement. Currently, the modeling of the VM generation behavior in the coking process primarily relies on computational fluid dynamics (CFD) simulation, which is computationally expensive and unsuitable for real-time applications. To address this issue, this study proposed a graph neural network (GNN)-based prediction method for the rate of VM generation in the coking process. First, a process dataset was constructed using CFD models, and temperature profiles were predicted using a deep neural network (DNN)-based feature augmentation method to improve feature representation in the VM generation rate prediction. Then, the graph data was structured based on the characteristics of the coking process, and the graph attention network (GAT) model was applied to predict the VM generation rate. Finally, peak internal gas pressure (IGP) was estimated using a mathematical model for IGP calculation, incorporating the GAT-predicted VM generation rate and the DNN-predicted temperature profiles. The results demonstrate that the proposed GAT-based method achieved superior prediction performance compared to other tested models, with the mean absolute error (MAE) value was 2.24×10 −4 , while the root mean square error (RMSE) value was 8.60×10 −4 , and coefficient of determination ( R 2 ) value was 0.993 for VM generation rate prediction. For peak IGP prediction, the lowest MAE value reached 0.604, and R 2 value was 0.846. The proposed model provides both theoretical and industrial significance, contributing to the digital transformation and optimization of the coking industry.