Accuracy and computational efficiency of suboptimal nonlinear predictive control based on neural models

This paper shows control accuracy and computational efficiency of suboptimal model predictive control (MPC) based on neural models. The algorithm uses on-line a neural model of the process to determine its local linear approximation and a nonlinear free trajectory. Unlike the fully-fledged nonlinear MPC technique, which hinges on non-convex optimisation, thanks to linearisation the suboptimal algorithm requires solving on-line only a quadratic optimisation problem. Two nonlinear processes are considered: a polymerisation reactor and a distillation column. In the first case MPC based on a linear model is unstable, in the second case it is slow. It is demonstrated that the suboptimal algorithm in comparison to the nonlinear MPC with full nonlinear optimisation: (a) results in similar closed-loop control performance and (b) significantly reduces the computational burden.