Prediction of NOx generation concentration based on feature mode decomposition integrated interpretative ensemble learning model for a 660  MW coal‐

Abstract

Accurately predicting NO _x concentration at selective catalysis reduction system is crucial for improving denitrification efficiency. However, combustion conditions change when the boiler is operating under deep peak regulation conditions, making it difficult to accurately test NO _x emissions in real time. The accurate NO _x concentrations were predicted using an optimized extreme gradient boosting (XGBoost) model combined with Shapley additive explanations (SHAP) analysis. The performance improvements facilitated by feature mode decomposition and fruit fly optimization algorithm demonstrate strong generality, as evidenced by reductions in RMSE of 20.2% and 4.2%, and reduction in MAPE of 35.9% and 26.8% in testing set. SHAP analysis indicates that the primary influences on the prediction results of the feature mode decomposition–fruit fly optimization algorithm–XGBoost model are over‐fired air, load, and O ₂ value, while primary air distribution and total air flow have lesser influences. Using a 350 MW boiler NO _x emission data to demonstrates the model's generalization capacity. The results show that 87.85% of the data points have relative errors less than 5%, and 70.56% of the data points have relative errors less than 0.5%. The model proposed in this work has achieved satisfactory results in prediction accuracy and training efficiency for NO _x concentration, and has balanced prediction accuracy with interpretability.