Automated learning and control using speedup and neural networks

Abstract

This paper studies complex dynamic neural network learning models. Backpropagation was used to train a neural network for dynamic simulation and control of a chemical stirred tank. The generalized delta rule algorithm was used to train the network minimizing the sum of squares of the residual. It was assumed that a historical database of plant inputs and outputs is available. A Pseudo Random Binary Sequence-PRBS was used as a disturbance. For training the database the 1% PRBS signal was superimposed upon its steady state value from SPEEDUP simulation. Once a trained neural network model was available, it then used in real time learning and pH control. The examined inverse and standard neural network models controllers achieved better performance than a conventional PI controller. Complex Internal Model Control - IMC achieved the best results in control and local stability. The obtained models in this paper improve noisy handling and reduce process variability. Some of these systems can be used for self - maintaining non-linear, multivariable models and day to day troubleshooting.