Modelling and Simulation for NOx Emission Concentration of SCR Denitrification System

Abstract

The selective catalytic reduction (SCR) denitrification system has the features of non-linearity, large lag and strong disturbance, when the operating condition changes. Based on mutual information (MI) and Kernel-based Orthogonal Projections to Latent Structures (KOPLS), the model for NOx emission concentration is proposed. The time-delay of each input variable is estimated by mutual information, and phase space construction is performed, KOPLS is utilized to modelling. KOPLS shows the merits of strong generalization, nonlinear fitting and anti-noise in the simulation of benchmark datasets. According to field data analysis, RMSE of MI-KOPLS in training and test are reduced by 17% and 22% respectively. Compared with KOPLS, MI-KOPLS predicts more accurately. Compared with other algorithms, RMSE and MAPE of MI-KOPLS reach minimum values 3.1886 mg/m3 and 13.5917% in test respectively, what indicates that the predicted value is the closest to real value, and the effectiveness of MI-KOPLS is verified.