Towards reliable and interpretable NOx emission prediction: A knowledge and data driven model for coal-fired boilers with spatiotemporal feature integration

Accurate prediction of nitrogen oxides (NOx) emissions is essential for controlling pollutant discharge and promoting the green, low-carbon transformation of coal-fired power plants. However, existing NOx prediction models exhibit notable limitations. They often fail to adequately capture the spatiotemporal dependencies arising from boiler load fluctuations, such as dynamic temperature field distributions in the combustion zone and the delayed migration effects of flue gas components. Moreover, conventional data-driven models lack integration with combustion mechanisms and domain knowledge, relying predominantly on input-output mappings. As a result, these models exhibit reduced predictive accuracy and limited interpretability, particularly under peak-shaving conditions. A Physically-Informed Graph Convolutional Network with Gated Recurrent Units (PI-GCN-GRU) model was proposed in this study to address the aforementioned challenges. A hybrid adjacency matrix was constructed by combining combustion process knowledge with the K-means clustering algorithm and was further integrated with time-series data to generate spatiotemporal graph representations. Furthermore, a physics-informed loss function was designed by embedding a dynamic NOx emission model as a physical constraint within the data-driven framework. The spatiotemporal graph data were input into the PI-GCN-GRU model, enabling it to simultaneously learn the spatial correlations and temporal dependencies among the operational features. The proposed model demonstrated superior performance on four datasets from a 1000 MW ultra-supercritical boiler, attaining a maximum prediction accuracy of R2 = 0.956 and achieving average MAE reductions of 51.7 % and 32.8 % over the conventional SVM and LSTM models, respectively. In addition, the SHapley Additive exPlanations (SHAP) method was applied to interpret the NOx prediction outcomes, offering scientific insights for combustion process optimization and pollutant emission reduction. The developed model was confirmed to deliver high stability and accuracy in predicting NOx concentrations, thereby providing a reliable solution for combustion system improvement and environmental compliance.