Improved Prediction of Nitrogen Oxides Using GRNN with K-Means Clustering and EDA

Abstract

The current study presented a generalized regression neural network (GRNN) based approach to predict nitrogen oxides (NOx) emitted from coal-fired boiler. A novel 'multiple' smoothing parameters, which is different from the standard algorithm in which only single smoothing parameter was adopted (Matlab neural network toolbox, for example), were assigned to GRNN model. K-means clustering algorithm was developed so as to reduce the number of smoothing parameters. The training data was firstly partitioned into groups (the number of groups was much smaller than that of training samples) using K-means clustering. A smoothing parameter was then assigned to this group. A recently emerging estimation of distribution algorithm (EDA) was employed to optimize the multiple smoothing parameters. EDA presented in this paper was a kind of optimization algorithm based on Gaussian probability distribution. As a case study, the proposed approach was applied to establish a non-linear model between the parameters of the coal-fired boiler and the NOx emissions. The results showed that the number of cluster has significant effect on the predictive accuracy of GRNN model. GRNN model with multiple smoothing parameters showed better agreement than that with only one smoothing parameter. The modeling errors on the testing subset were 1.24% and 1.62% for GRNN models trained by the present algorithm and the standard algorithm, respectively.