Abstract: Selective catalytic reduction is a dynamic process with dynamic time delay. A modeling algorithm based on dynamic time delay analysis and typical sample selection is proposed to establish an accurate prediction model of NO_x emission concentration of SCR outlet. First, through the mechanism analysis and Pearson correlation analysis of actual production data, the input variables of the model are determined. Secondly, the sliding time window strategy is adopted to compute the mutual information entropy of variables within the sliding time window. And the dynamic delay time of each variable for NO_x emission concentration is obtained based on the mutual information entropy. Thirdly, in order to reduce the number of training samples, the K-Means clustering algorithm is employed to select the sample set after time delay analysis. The cluster centers with representativeness and diversity are screened out. Finally, an extreme learning machine algorithm is designed to build a dynamic prediction model of NO_x of SCR outlet. Experimental results based on operation data of a 1000MW ultra-supercritical boiler show that the prediction errors of the established models are less than 5%, which could accurately predict the NO_x emission concentration at the outlet of SCR system.