Accurately predicting NO_x emission is conducive to reducing the cost of SCR flue gasdenitration and optimizing the coal combustion process. This paper uses partial least squares(PLS) method to carry out variable importance projection(V_(ip)) analysis on the raw data of coal-fired boilers. Based on V_(ip)

the raw variables are ranked

and the optimal 20 variable subsets are used as the input of deep belief networks(DBN)

and finally the PLS-DBN model for NO_x emission prediction is obtained. The PLS-DBN model and the least square supportive vector model(LSSVM)

DBN and error back propagation neural network(BPNN) model are applied in 3 000 groups of training set and 500 groups of test sets of a 660 MW unit boiler. The results show that

the prediction error of training set and test set of the PLS-DBN model is both small

and the prediction accuracy of root-mean-square error of not more than 2% on the training set and test set is 0.940 and 0.714

respectively

which is the highest

indicating the PLSDBN model has higher prediction accuracy and stronger model generalization ability than other three conventional NO_x prediction models.