Neuro-genetic PID autotuning: time invariant case Original Research Article

The Proportional, Integral and Derivative (PID) controllers are widely used in industrial applications. Their popularity comes from their robust performance and also from their functional simplicity. Because the plant to be controlled is time varying, or due to components ageing, or yet because of changes in process dynamics due to alterations of operation conditions, these controllers need to be regularly retuned. Since an accurate tuning is a time-consuming operation, and as even a single plant can have several of these small controllers, methods that automate the tuning process are economically important. In this paper a recent approach for PID autotuning, involving neural networks, is further extended, to incorporate multiple tuning criteria, and to make use of on-line experimental data. In this paper neural network models of tuning criteria, together with the use of genetic algorithms (GA), are proposed to achieve this aim. Simulation results show that, for the case of time-invariant plants, trained multilayer perceptrons are good models and generalise well. The closed loop unit step response obtained with the neuro-genetic approach compares favourably with the one achieved using a standard gradient technique with dynamic closed-loop simulation. More important, the proposed approach takes a fraction of the time spent by the standard technique, without the need of perturbing the closed-loop system.