Nonlinear System Modeling Using Multiple Neural Networks

Abstract

Improved predictions can be obtained by using multiple networks instead of a single optimal network as usual. Principal components regression is used to obtain stacking weights of multiple networks other than the least square method. Model accuracy and robustness can be significantly improved by using multiple neural networks (MNN). The proposed method has been applied and evaluated for three examples, including a steadystate nonlinear model, a dynamic nonlinear model and dynamic modeling of a practical depropane distillation column. Results obtained demonstrate that this approach can improve the performance of neural network based nonlinear models.