Hybrid Model Based Optimal Control for a Metallurgy Process Z.F.Qiu* , **. G. Deconinck* W.H.Gui**. C.H.Yang** *Katholieke Universiteit Leuven, Leuven, Belgium (Tel: 032-16321810; e-mail: zhifeng.qiu@ esat.kuleuven.be). **Central South University, Changsha, China Abstract: This paper applies hybrid modeling method based optimal control in industrial process. Hybrid modeling method combines a priori information with a nonlinear residual compensation technique to build a global model which predicts alumina raw pulp slurry quality. Process control is accomplished based on blending expert knowledge with multi-objective hierarchy reasoning approach. Through the coordination of model and controller, the optimal control of blending process is achieved. Application results show that the proposed method can resolve optimization problems of a kind of industrial processes characterized by time delay and multi-constraints. 1. INTRODUCTION The increased fierce competition over the last few years combined with large alumina price fluctuations has forced most alumina and metallurgy companies to find ways to streamline and optimize plant operations. At the same time, increased computing power combined with improved modeling tools has provided the opportunity to use different modeling technology for generating accurate product quality predictions. Moreover, modeling such complex industrial processes is a complicated procedure and is traditionally done using white box modelling or black box identification (Sjőberg et al., 1995), where white box modelling means that the model is constructed using scientific relations that completely describe the process. The black box identification of making a process model is done by using a standard parametric model is adapted to measured data obtained from the process. Many white box modelling techniques, so-called first principle, exist for metallurgy processes, such as mass and energy balance, physicochemical reaction and thermodynamic mechanic. However it’s very hard to deduce a precise physical formula, no matter its structure or parameter, due to not only the inherent complexity of such process but also the practical production environment. On the other hand, most model-based control strategies (Rohit, 2007; Zhang, 2007) make use of general linear or nonlinear black- box techniques to model the relationship between process input and output variables. This kind of method has the advantage of bypassing the complexity and the uncertainty of the physical systems. However, such models, especially nonlinear ones, may become themselves complex and involve a large number of parameters. Searching for the desired model parameters in a high dimensional model parameter space is prone to local minima and could lead to an inappropriate model. So, a new strategy combining these first principles with general black-box techniques is adopted as a means to trade-off between the complexity and the performance of such complex industrial processes at the supervisory control level. Generally, in hybrid modeling approach (Bohlin, 2001; Sohlberg, 2003), fundamental knowledge captured from industry process is used to define a prior parametric model with fixed structure derived from either first principle, existing empirical correlation or mathematical transformation, while the unknown part is modeled by a black-box model. Modeling this unknown part is usually much less complicated than modeling the whole process using a black-box model. The main contribution of hybrid modeling approach (Chen et al., 2004; Li et al., 2004) is as follows. First, this kind of model contains more physical meaning than a total black-box model thanks to the introduction of mechanical knowledge about the process. Second, some inherent problem of black- box technology can be overcame, for instance, the dimension and the feasible region of the parameter space are reduced that can hopefully overcome some identifiability problems, at the same time, the model has better generalization capability than a complete black-box model when the hybrid model structure is appropriate. Last but not least, the whole hybrid model is more suitable to be used in the practical engineering. In this paper, as the key to the process control success, the hybrid modelling technology is realized successfully in a metallurgy process, alumina blending process control. The control system is composed of a prediction model and an expert controller. The prediction model is employed to forecast raw pulp slurry quality, which consist of mass balance equation and neural network model as estimator for some of the important process parameters as well as compensator of the physical model. The function of controller is to implement optimal raw material ratio setting control, which is based on expert knowledge. Through the coordination between the prediction model and expert Proceedings of the 17th World Congress The International Federation of Automatic Control Seoul, Korea, July 6-11, 2008 978-1-1234-7890-2/08/$20.00 © 2008 IFAC 10844 10.3182/20080706-5-KR-1001.2638