Modelling case study for holistic production control D.Gradiˇ sar * M.Glavan * J. Kocijan *,** S. Strmˇ cnik * * Joˇ zef Stefan Institute, Ljubljana, Slovenia (e-mail: dejan.gradisar@ijs.si). ** University of Nova Gorica, Nova Gorica, Slovenia Abstract: The paper describes the part of research based on the idea of a holistic production control and optimisation concept which would be based on a simple model of only a few principal production Performance Indicators (pPI) (such as production rate, quality of products and associated costs). The model would be extracted from historical plant data using various statistical, identification and data mining techniques. Such a concept would enable simple upgrading of existing MES and similar tools which enable collection of a vast amount of data, but are relatively weak in offering support for decisions and control. The paper presents the modelling part of the procedure for the design of the holistic control and optimisation of the production, i.e. the derivation of a simple pPI model using neural networks. The modelling is demonstrated on the well known Tennessee Eastman benchmark process, which represents a plant-wide industrial production of chemical components. Keywords: Holistic Production Control, Production Performance Indicators, Black-box modelling. 1. INTRODUCTION The modern business environment demands an instant re- sponse to customers needs, a short product life cycle, prod- uct diversification, minimal inventories, extremely short lead times, etc. Advanced manufacturing requires quick and accurate decisions and actions at all management levels in a factory. The demand for high cost effectiveness has turned modern industry away from the planned pro- duction concept to an order-driven one. This has entailed a new concept of management based on an online estimation of the current situation together with efficient decision making and execution. There exists different tools that help managing activities within that kind of business envi- ronment, however we can notice some lack of capabilities which could improve the performance at the production control level. Over the years, the area of plant-wide control has attracted the process engineering community and a number of solu- tions were suggested. On the one hand research is dealing with a hierarchical decomposition of the original control design problem based on heuristic rules. The heuristic logic is developed so to keep process variability and therefore the operational plant objectives under acceptable limits for a given set of disturbances (see Luyben et al. [2004] or Skogestad [2002]). In recent years closed-loop control is being used also at the so called production control level – the level between busi- ness and process control levels. One promising approach here is Real-Time Optimisation (RTO). Engell [2007] de- fines RTO as a model based upper-level control system that is operated in closed loop and provides set-points to the lower-level control systems in order to maintain the process operation as close as possible to the economic optimum. Summary of recent developments and applica- tions of dynamic real-time optimisation is given by Kadam and Marquardt [2007]. An aspects of how to use Model Predictive Control within the RTO structure is presented by Rawlings and Amrit [2009]. On the other hand, there are mathematically oriented approaches based on the solution of a given large scale mixed integer nonlinear programming dynamic optimisa- tion problem. In the limit those solution methods should be able to simultaneously determine the optimal process units size and their interconnections as well as the optimal control scheme configuration (see Biegler and Grossmann [2004] for a review). The main idea of our research is to derive a holistic control and optimisation concept which would be based on a simple model of only a few principal production Perfor- mance Indicators (pPI) (such as production rate, quality of products and associated costs). The model would be ex- tracted from historical plant data using various statistical, identification and data mining techniques. Such a concept would enable simple upgrading of existing MES and simi- lar tools which enable collection of a vast amount of data, but are relatively weak in offering support for decisions and control. The control and optimisation system based on the proposed approach would thus increase flexibility and performance of production control, simultaneously achieving the most important global production objectives and to certain extent eliminating the subjective decisions of production managers. 2010 Management and Control of Production Logistics University of Coimbra, Portugal September 8-10, 2010 978-3-902661-81-4/10/$20.00 © 2010 IFAC 186 10.3182/20100908-3-PT-3007.00035