Neural Predictive Control for Wide Range of Process Systems

In this paper a Neural Predictive Controller (NPC) designed to control a broad class of process systems. Neural network identification yields nonlinear global model of the unknown system. LevenbergMarquardt (L-M) optimization method is used to find optimal control signal to minimize future errors of the objective function of predictive controller. Inequality constraints of actuators are added to the objective function through a penalty term which increases drastically as it approaches the limitations. To use the controller for wide range of process systems, an initial phase runs before the main controller to determine parameters. This phase moves the system output to operating point and applies PID controller with APRBS reference signal. The gathered data are used to estimate parameters such as pure delay, prediction horizon, control coefficient and identification order. To validate the approaches, the controller has implemented in level, pressure and flow pilot plants and compared with conventional controller which shows faster and smoother tracking results.