Abstract In recent years, multilayer feedforward neural networks have been used for chemical process identification [9]. One argument used repeatedly against neural network models is that they are “black-box” input/output models. The exact nature of the mapping from inputs to outputs is not easily understood. However, the advantage of neural network models is that they are able to predict process behavior accurately without any a priori knowledge of the process. Alternatives to input/output models are first principles (FP) models. The FP models describe the details of a process, but they are hard to develop, complex and difficult to solve. This paper discusses an approach to integrating FP models and neural network models, Integrated Neural Network (INN) models. Model mismatch between the FP model and the process is captured by the neural network. A polymer reactor system is chosen as a case study to demonstrate the technique.
[1]
Kurt Hornik,et al.
Multilayer feedforward networks are universal approximators
,
1989,
Neural Networks.
[2]
Hong-Te Su,et al.
Identification of Chemical Processes using Recurrent Networks
,
1991,
1991 American Control Conference.
[3]
C. R. Cutler,et al.
Dynamic matrix control¿A computer control algorithm
,
1979
.
[4]
Thomas J. McAvoy,et al.
Nonlinear PLS Modeling Using Neural Networks
,
1992
.
[5]
W. Harmon Ray,et al.
On the Mathematical Modeling of Polymerization Reactors
,
1972
.
[6]
F. Girosi,et al.
Networks for approximation and learning
,
1990,
Proc. IEEE.
[7]
Mark A. Kramer,et al.
Improvement of the backpropagation algorithm for training neural networks
,
1990
.