In view of the complex variation law and strong autocorrelation of nitrogen oxides emission mass concentration of circulating fluidized bed (CFB) boiler, by using relevant variables and their historical information, ensemble learning online models of nitrogen oxides emission mass concentration are established. The ensemble learning online models include the autoregressive integrated moving average (ARIMA), random forest (RF), gradient boosting (GBDT), and eXtreme gradient boosting (XGBoost) model. The prediction results are compared and selected, among which the GBDT regressor is the best. In order to further improve the prediction effect of the model, a GBDT differential regression model is established by combining the first-order difference with the GBDT regression algorithm. The tests show that the established GBDT differential regression model has better prediction performance than the aforementioned models. The mean squared error of the predicted value is 20.2% lower than that of the simple GBDT regressor, and 46.5% lower than that of the online sequential extreme learning machine (OS-ELM) model used in the reference. The online model also fully considers avoiding the influence of the instrument purge process, and has strong practicability.