The coking industry is facing increasing demands for improved energy efficiency, reduced carbon emissions, stable coke quality, and more intelligent operation. However, the coal coking process is characterized by strong nonlinearity, significant time delays, complex thermochemical interactions, and limited process observability, which pose substantial challenges to real-time prediction, autonomous control, and reliable operation. To address these issues, this study develops an artificial intelligence-enabled digital twin framework integrating physics-based simulation, data-driven modeling, intelligent control, and process monitoring for coal coking.
A computational fluid dynamics model of an industrial coke oven was first established to describe the transient heat-transfer behavior and generate high-fidelity data for data-driven modeling. Machine learning models, including neural-network and ensemble-learning approaches, were subsequently developed to achieve rapid prediction of the coking heat-transfer process. Intelligent optimization algorithms were introduced to improve model performance, while SHapley Additive exPlanations (SHAP) were employed to enhance model interpretability. Experimental data from a laboratory-scale coke oven were further used to evaluate the generalization capability of the developed models.
To characterize volatile matter evolution during coking, a graph-based modeling strategy was proposed by incorporating temperature information into a graph-structured representation of the process. A graph attention network was then used to capture the spatial and temporal dependencies associated with volatile matter generation and internal gas pressure evolution. The results demonstrate that graph-based learning provides an effective approach for representing the strong coupling between local thermal conditions and volatile release behavior.
For intelligent process regulation, surrogate models were constructed to reproduce the dynamic response of the coking process and were integrated with deep reinforcement learning for closed-loop control. The developed control framework enables dynamic adjustment of heat-transfer and volatile-matter-generation behavior under changing operating conditions. The combination of reinforcement learning with conventional control was also investigated to improve response performance and disturbance rejection, while maintaining adaptability to complex dynamic operating scenarios.
In addition, a Transformer-based monitoring framework was developed for fault detection and fault identification. An unsupervised Transformer autoencoder combined with clustering was employed to detect abnormal process states, while a supervised Transformer classifier was used to identify different fault types. The proposed monitoring approach showed stable performance under different coal-blending schemes and operating conditions, demonstrating its potential for robust process supervision.
Overall, this study integrates mechanistic simulation, machine learning, graph-based modeling, reinforcement learning, and intelligent fault monitoring within a unified digital twin framework. The proposed approach provides a systematic pathway toward rapid process prediction, autonomous regulation, and reliable operation, and offers methodological support for the development of intelligent coke-oven operation systems.
 

coal coking,digital twin,machine learning,deep reinforcement learning,fault diagnosis