Process Modelling and Control with Neural Networks: Present Status and Future Directions

Abstract

Even though there exist numerous traditional approaches for solving nonlinear control tasks their realization in practice often proves to be difficult, mainly because of i) insufficient analytical knowledge of the system to be controlled, and ii) because of incomplete knowledge of the physical parameters of the system.

Neural networks can cope with these problems because of their capability to realize multivariate, nonlinear transformations and to learn these merely by empirical information, i.e. by representative training examples (data driven modeling).

In this contribution we discuss neural methods for modeling and control and illustrate their use in real world applications. In order to improve the corresponding conventional solutions, it is crucial to i) incorporate prior knowledge (analytical and/or rule based) into the neural network and ii) to optimize its architecture by refined learning methods. We point out the future need for neural control of the complete process rather than controlling its parts independently.

Further, we discuss the necessity to combine neural methods with the modern techniques of Nonlinear Dynamics for the modeling and control of processes with highly nonlinear dynamic behavior.