Abstract
The objective of this research is to develop a nonlinear empirical model structure and an associated parameter estimation algorithm based on artificial neural networks (ANNs), and to further use it for the identification of a highly nonlinear process system component, namely a U-Tube Steam Generator (UTSG). The proposed model structure is called a Recurrent Multilayer Perceptron (RMLP). RMLP is a hybrid feedforward and feedback neural network, which in addition to the feedforward connections it exhibits local information feedback, through time delayed recurrency and cross-talk. A static and a dynamic learning algorithm is derived for parameter estimation (learning), and both algorithms are used to train the RMLPs. The capability of the RMLP to identify nonlinear systems using dynamic learning is demonstrated through several examples. Traditional and ANN based model structures are compared for their effectiveness to identify nonlinear systems. Comparisons of the chosen model structures is accomplished through a number of deterministic and stochastic examples. The responses of the identified models to different test signals, unknown during estimation, are presented for investigating their predictive performance. As expected, nonlinear model structures perform better than their linear counterparts. Furthermore, among the nonlinear structures, the RMLP based models exhibit improved predictive performance when identifying stochastic systems. For deterministic systems identification, however, feedforward multilayer perceptron (FMLP) and RMLP based empirical models reveal comparable accuracy. The effectiveness of the RMLP nonlinear empirical model structure with static learning is further demonstrated by developing two models for a UTSG, each valid in the vicinity of an operating power level. A significant drawback of the static learning algorithm has been the excessively long off-line training times required for the development of an even simplified model for a UTSG, hindering the further development of a single model valid in the entire UTSG normal operating envelope...
Chong, Kil To (1993). Nonlinear dynamic system identification using recurrent neural networks. Texas A&M University. Texas A&M University. Libraries. Available electronically from
https : / /hdl .handle .net /1969 .1 /DISSERTATIONS -1462059.