A digital twin of coal-fired boiler for physical fields prediction using flow-field-informed POD-XGBoost reduced-order modeling

The online accurate prediction of temperature and species in coal-fired boilers with high spatial resolution is crucial and remains challenging for combustion monitoring and optimization. Among the prediction methods, the approach that integrates proper orthogonal decomposition (POD) with surrogate models has emerged as a promising approach in recent years, delivering temperature and species with high spatial resolution. However, this method relies solely on boundary conditions and lacks detailed local flow field information for predictions, leading to significant errors in regions with intense turbulence. To address this issue, this study developed a combustion digital twin using flow-field-informed POD-XGBoost reduced-order modeling by incorporating computationally efficient non-combustion CFD flow information, namely velocity magnitude, as additional input to the surrogate models for more accurate real-time predictions. The results show that the established digital twin outperforms the parameter-based predictions for all three physical fields overall. The median normalized root mean squared errors (NRMSE) for temperature, O2, and CO are 0.0240, 0.0333, and 0.0442, respectively. The flow-field-informed digital twin also exhibits high computational efficiency, requiring only 5 % of the computational time of traditional CFD. The flow-field-informed digital twin exhibits promising potential for combustion monitoring and optimization in coal-fired boilers.