Abstract: To address the issues of large time delays, complex chemical reactions, and multiple influencing factors on the outlet NOx concentration in the selective non-catalytic reduction (SNCR) denitrification system for waste incinerators, a dynamic modeling method based on maximum information coefficient (MIC) and bi-directional long short-term memory network was proposed. Firstly, by analyzing the SNCR denitrification mechanism and influencing factors, the variables related to outlet NOx were preliminarily selected. Then, based on the MIC algorithm, variables with strong correlation with outlet NOx were selected and redundant variables were removed. Next, the sliding window method and MIC algorithm were used for data delay estimation to complete the data reconstruction. Finally, a dynamic model of the SNCR denitrification system was built based on the bi-directional long short-term memory deep learning algorithm.The results showed that after variable screening and delay estimation, the accuracy of the dynamic model was significantly improved, and the average absolute percentage error of the model was approximately 8.62%, which decreased by 18.6%, 18.1% and 12.5% compared with the BPNN, LSTM and GRU models, respectively. Therefore, this modeling method has higher precision and better fitting effect, which can be more effectively applied to the actual field.

Key words: selective non-catalytic reduction, maximum information coefficient, bi-directional long short-term memory network； NOx； dynamic modeling