Suboptimal nonlinear predictive control based on multivariable neural Hammerstein models

Abstract

This paper describes a computationally efficient nonlinear Model Predictive Control (MPC) algorithm in which the neural Hammerstein model is used. The MultipleInput Multiple-Output (MIMO) dynamic model contains a neural steady-state nonlinear part in series with a linear dynamic part. The model is linearized on-line, as a result the MPC algorithm requires solving a quadratic programming problem, the necessity of nonlinear optimization is avoided. A neutralization process is considered to discuss properties of neural Hammerstein models and to show advantages of the described MPC algorithm. In practice, the algorithm gives control performance similar to that obtained in nonlinear MPC, which hinges on non-convex optimization.

On-line set-point optimisation and predictive control using neural Hammerstein models

“…Two types of neural Hammerstein models of multivariable processes are thoroughly discussed in [28]. In the first case the nonlinear steady-state part of the model is realised by only one neural network, in the second case as many as n x neural networks are used.…”

“…Multivariable Hammerstein models discussed elsewhere, e.g. in [20,28], do not take into account disturbance signals explicitly, which are essential for set-point optimisation.…”

“…The MPC algorithm considered here is an extension of the algorithm described in [28] because disturbances e are taken into account.…”

“…Hammerstein models can be efficiently used for modelling of various chemical processes: e.g. distillation columns [10,12,23], heat exchangers [10], neutralisation reactors [12,22,28,42]. An excellent review of identification algorithms and applications of Hammerstein models in process modelling, control and fault diagnosis is given in [20].…”

“…In particular, Hammerstein models can be used for prediction in MPC algorithms, e.g. [6,12,16,28,42].…”