Study On Nonlinear Model Predictive Control Algorithm Based On Combination Model And Its Application

It's common that many of nonlinear systems in chemical processes can be represented by a combination model which consisting of a static nonlinear part and a linear dynamic part. The dissertation studies these systems and proposes two types of predictive models and three types of Nonlinear Model Predictive Control (NMPC) algorithms. The proposed algorithms are apllied to the liquid level control of a Coupled Water Tanks system, which verify their effectiveness. The main contents are as follows:1) After analyzing and comparing with the predictive model of NMPC algorithms mentioned in literatures, a numerical steady-state model is brought forward, which is a map of a set of values representing the steady-state relation between system input and output. Furthermore, two type of predictive models, numerical steady-state ARX combination model and BP (back propagation) neural network ARX combination model, are proposed.2) The nonlinear steady-state characteristic of SISO (Single-Input Single-Output) nonlinear system is described by a numerical steady-state model rather than an analytic model, thus simplifing the procedure of model identification. With the numerical steady-state ARX combination model which takes a similar form as those of Hammerstein model where a static nonlinear part preceding a linear dynamic part, the parameters of the off-line identified ARX model are modulated on-line according to the steady-state model. Thus a parameter varying ARX model is used to implement NMPC, which can be named as numerical NMPC algorithm.3) A new combination model in which the numerical steady-state model being parallel to the linear ARX model is proposed to represent SISO system. With the parameters of the offline identified ARX model are mainly modulated on-line according to the steady-state model, the recursive least squares method (RLS) which is commonly used in adaptive control is employed to improve the self correcting ability of the predictive model. Therefore, the steady state performance of the system when closing to operation points is also improved together with the performance of fast response to setpoints change. Online optimization is solved by sequential quadratic programming. Adaptive NMPC algorithm which can deal with system possessing strong nonlinear characteristic is proposed.4) The nonlinear steady-state characteristic of MIMO (Multi-Input Multi-Output) nonlinear system is described by a BP neural network and a BP-ARX combination model is proposed. The parameters of the off-line identified ARX model are modulated on-line according to the BP steady-state model, and the NMPC is carried out by a parameter varying ARX model. Meanwhile, based on the above mentioned adaptive NMPC algorithm, a shift coefficientξis defined to constrain the effect time of adaptive algorithm. If and only if the errors between system real measurement outputs and output setpoints less than the set value ofξthe RLS algorithm comes into effect. A parameter varying NMPC (PV-NMPC) algorithm is proposed finally.5) The architecture and functional modules of NPC (nonlinear predictive control) software which are based on the algorithms proposed in the dissertation are discussed.6) The algorithms described in chapter 3 and chapter 4 are tested on a Coupled Water Tanks system. The control results are compared with PI plus feedforward controller.