NONLINEAR MODEL BASED PREDICTIVE CONTROLLER USING A FUZZY-NEURAL WIENER-HAMMERSTEIN MODEL Yancho Todorov * , Margarita Terziyska ** , Michail Petrov ** * Institute of Cryobiology and Food Technologies 53, Cherni Vrah blvd., 1407 Sofia, Bulgaria fax: +359 2 868 33 73 .and e-mail: yancho.todorov@ieee.org ** Technical University- Sofia, branch Plovdiv 25, Tzanko Dustabanov str., 4000, Plovdiv, Bulgaria fax: +359 32 659 530 .and e-mail: mng@tu-plovdiv.bg , mpetrov@tu-plovdiv.bg Abstract: It is presented in this paper a method for designing a nonlinear model predictive controller. The controller is based on a hybrid Wiener-Hammerstein fuzzy-neural predictive model and  simplified gradient optimization algorithm. The proposed approach is used to control the product temperature in a Lyophlization plant. The controller efficiency is tested and proved by simulation experiments in Matlab & Simulink. Keywords: Model Predictive Control, Fuzzy models, Lyophlization 1 INTRODUCTION The underlying philosophy of Model Predictive Control (MPC) consists of minimization of a performance objective function with respect to future input moves, over a finite time horizon. Simplicity of design combined with ability to tackle realities and interactions has helped MPC to achieve its current popularity in the process industry (Patwardahan et al. 1998). In MPC a process dynamic model is used to predict future outputs over a prescribed period called predic- tion horizon. Afterward, the model outputs are used to compute the future control actions by minimizing the predefined cost function. Most technological processes are inherently nonlinear. Under this circumstance, together with higher product quality specifications and increasing productivity demands require operating systems closer to the boundary of the admissible operating region. Linear models are often inadequate to describe the process dynamics and nonlinear models have to be used in these cases. Several excellent reviews of the main NMPC principles and advantages/disadvantages of can be found in (Rawlings et al. 2000), (Morari et al. (1999), (Findeisen et al. 2002). Researchers have proposed several ways to equip MPC with the capability to deal with nonlinear processes. The approaches to Nonlinear Model Predictive Control (NMPC) can be categorized into two groups: those based on first principles models of the process and those based on black-box models identified from input-output data. The nonlinear process models based on Artificial Neural Networks and Fuzzy Logic belong to the second group of black-box models. Typical for them is the precise description of the plant process by a set of linear submodels. In this way the design of a model predictive controller can be gratefully simplified. Also, with the fuzzy-neural models it is possible to cover a broad range of the process operating conditions. One of the most frequently studied class of nonlinear models are the so-called block oriented nonlinear models, (Abonyi et al. 2000) which consist of the interconnection of Linear Time Invariant (LTI) systems and static nonlinearities. Within this class, two of the more common model structures are the Hammerstein and the Winner models.