Pergamon Computer~ chem. Engng, Vol. 21, Suppl., pp. S1093-S1098, 1997 © 1997 Elsevier Science Ltd All fights reserved Printed in Great Britain PII:S0098-1354(97)00195-6 0098-1354/97 $17.00+0.00 A Practical Method of Removing ill- conditioning in Industrial Constrained Predictive Control Akio Ishikawa* Masahiro Ohshima** and Masataka Tanigaki** * Process Technologyand Engineering Center, Mitsubishi ChemicalCorporation, Kurosaki Plant, Kitakyusyu 806, JAPAN ** Department of Chemical Engineering, Kyoto University,Kyoto 606-01, JAPAN ABSTRACT - At any control execution of large scale DynamicMatrix Control [DMC] TM, the process may require excessive input movements. This problem arises from the ill-conditioned internal modelsof the predictive controller. In this paper, a practical method of removing the ill-conditioning is developed by making the best use of the Singular Value Analysis (SVA) and the characteristic of LP optimizer in DMC. The proposed method is applied to an industrialammoniaplant, where non-square (25 controlled- 12 manipulated variables)DMC controller is installed. By redesigningsome internal models which are indicatedby the method, the resulting controller can suppress the excessive input movementsand improve the control performance. INTRODUCTION In the last decade in Japan, Model Predictive Control (MPC) has been employed as an advanced process control technique at many factories and has been widely applied to large industrial processes (Ohshima, et. al., 1995). Mitsubishi Chemical has also implemented one of the MPC schemes, Dynamic Matrix Control (DMC), on six plants and achieved significant economic benefits. Even though the benefits were higher than expected, some control problems remain to be solved for further improvement of the controllability. For example, in the application of DMC to a large scale process, a large number of dynamic process models, such as finite response models, are prepared through several identification tests in order to construct the internal models used for output prediction. Some models show a small steady-state gain. Some of the process models might be negligible for designing the internal models of the predictive controller, but some are not. At the identification stage, it was difficult to know whether the control performance is deteriorated by neglecting such small gain models. The preparation of the internal models depends mostly on operational knowledge as well as experiences. Neglecting one model sometimes causes the excessive input movement at some particular control execution and may end up unstable control system. Singular value analysis (SVA) is often employed to check the ill-conditioning in the input-output paring (Grosdidier, et. al., 1985; Skogestad, et. al., 1987) and the dynamic matrix appearing in the least-squared input calculation (Seborg, et. al., 1988; Wilkinson, et. al., 1994). However, no SVA application has been reported on the DMC with Linear Programming (LP) optimizer controlling non-square processes. The DMC with LP optimizer, which is called DMC- LP controller hereafter, uses a Linear Programming to perform the local steady-state optimization subject to hard input and soft output constraints. As proved later in this paper, at every control execution, the LP works as if for selecting an equal number of the controlled and the manipulated variables from the full scale non-square process, and controlling the selected square process. This paper proposes a method of evaluating the ill- conditioning in the non-square process controlled by the DMC-LP controller. In the method, the SVA is employed to calculate a condition number and evaluate ill-conditioningof the selected square process appearing in LP executions. Then, the internal models are redesigned so as to make the condition number smaller and to remove the ill-conditioning. The proposed method is applied to one of the processes in Mitsubishi Chemicals to show its effectiveness. CONTROL SYSTEM CONFIGURATION To perform the optimal operation on the large scale $1093