Nonlinear predictive control of temperature in long duct using specially designed neural model

Abstract

This paper describes a nonlinear Model Predictive Control (MPC) algorithm for a distributed parameter thermal system (a long duct). For prediction a specially designed neural model of the process is used. The model consists of a set of local neural sub-models, which calculate temperatures for a number of predefined locations of sensors, and a neural interpolator, which calculates the temperature for any sensor location. In order to obtain a computationally simple MPC scheme, the predicted output trajectory of the process is linearised on-line which leads to a quadratic optimisation MPC problem. It is shown that due to nonlinearity of the process, the classical MPC algorithm based on linear models is unable to give satisfactory control quality whereas the described nonlinear MPC algorithm leads to good control performance. The paper also studies the effect of model pruning (removing some of the sub-models) on the performance of MPC.