NOx emissions prediction based on mutual information and back propagation neural network using correlation quantitative analysis

Abstract

Information on Nitrogen oxide (NOx) concentrations play a significant role in aftertreatment systems. In this work, a method to estimate NOx emissions by means of mutual information (MI) and back propagation neural network (BPNN) was introduced. All measured signals were ranked by MI value and the most significant parameters were classified according to their physical meanings. The model inputs were selected by analysis of the classified groups. The forecasting model was developed by the BPNN algorithm to predict raw NOx emissions and NO mass flow rate (MFR) before Selective catalytic reduction (SCR) with selected input variables. Ranking, classification, selection, and training were carried out under steady-state conditions and the world harmonized stationary cycle. The verified BPNN network could well predict raw NOx emissions and NO MFR before SCR. Compared to static map prediction, the mean absolute deviation and root mean square error of BPNN are reduced by about 15%, which also indicated that the MI-based feature selection method was effective. The proposed approach is a generic approach for NOx emission prediction, which could also reduce the requirement for expert knowledge on feature selection, has a lower computational cost, and could be used in engine and aftertreatment control system of real driving vehicle.