An experimental study of controllers based on neuronal networks

• Javier Bastidas Universidad del Zulia-Venezuela
• Carlos Vinante Universidad del Zulia-Venezuela

Abstract

In this paper a predictive controller is presented, the output of which is obtained at each sampling time by numerically minimizing the merit function real error between the desired value and the output produced by the neural network that models the process dynamics. The on-line solution to this non-lineal problem at each sampling moment corresponds to the output obtained using the direct method and the dynamic of the realizable inverse process. In comparison with the topology of the classic Internal Model Controller (IMC), this controller uses only the direct neural network and does not have the error problems of a stationary state due to the inexactness of the direct and inverse neuronal models. In order to demonstrate an experimental application of this controller, the proposed algorithm is applied to control the flow in experimental equipment used to calibrate valves. This is an interesting process because it includes non-linear considerations and dead time. The experimental results indicate that, for processes with non-linear gains and time delay, this algorithm performs better than the controllers based on linear process models.