Data Driven Modeling Using an Optimal Principle Component Analysis Based Neural Network and Its Application to a Nonlinear Coke Furnace

Abstract

In order to fully exploit the data information among process variables, an optimization based principle component analysis (PCA) using a neural network is proposed. First, a new RV similarity criterion of the PCA variable selection method is developed to select the main variables and construct the nonlinear industrial process. Second, a radial basis function neural network (RBFNN) is utilized to construct the nonlinear process model, where the modeling accuracy and RV criterion are optimized by an improved multiobjective evolutionary algorithm, namely, NSGA-II. To obtain the optimization of the structure and parameter of the RBFNN, encoding, prolong, and pruning operators are designed. The RBFNN with good generalization capability will then be obtained based on root mean squared error of the training and testing data considering the Pareto optimal solutions. The proposed approach has efficiently selected the main disturbance of the chamber pressure control loop in a coke furnace, and the RBFNN has obtained satisfactory data extraction accuracy compared with the other three typical methods.