Abstract: In order to better control the environmental emissions of circulating fluidized bed boilers, aiming at the complex characteristics of the variation of nitrogen oxide emission concentration, the ensemble learning regression algorithms such as Random Forest (RF), Gradient Boosting (GBDT), eXtreme Gradient Boosting (XGBoost) were used to establish the offline prediction models of nitrogen oxide emission concentration, and the prediction results were compared and selected. The test results showed that the ensemble learning regression models has obvious advantages over the traditional linear regression, and the GBDT regressor was the best. In order to further improve the prediction effect of the models, it was proposed to combine the XGBoost classifier with GBDT and XGBoost regressors to form an ensemble learning classification comprehensive model. The test results showed that the comprehensive model has better prediction performance than the single integrated learning regression model, and the Mean Squared Error (MSE) of the predicted value was 1.9% lower than that of the single GBDT regressor model. The combination method used in the comprehensive model was compared with the stacking generalization combination method used in the references. The test results showed that the MSE of the predicted value of the comprehensive model was 2.6% lower than that of the stacking generalization model. The model was further applied to analyze the influencing factors of nitrogen oxide emission concentration in detail, and the optimal quantitative operation interval was given, which improves the operator's previous fuzzy qualitative understanding.