Interesting Approaches in Model Construction for Controlling a Glass Melt

In this year’s competition of the international scientific association, EUNITE, three teams have succeeded in predicting the temporal development of the five response signals in the process control of a glass melting tank (with 29 input parameters) over a period of two weeks. The – weighted – deviation of the modelled values from the real values was 0.3 per cent. The winners will present their results at the EUNITE Conference from 10th to 12th July in Oulu, Finland. As an expression of its gratitude, Schott Glas will award prizes in the amounts of 5000 € (1st place), 3000 € (2nd place), and 1000 € (3rd place). “We could not actually apply the concrete values from the prediction for plant control, perhaps because some decisive parameters had not yet been recorded, but the proposed model approaches are of great value to us”, said Dr. Katharina Lankers, who had arranged the competition on behalf of Schott Glas.

EUNITE is a European association of scientists at universities and in industry; it is supported by the European Union with the objective of forming a network for excellence. EUNITE is dedicated to improvements in so-called intelligent, adaptable systems. The modelling abilities of the scientists are tested in an annual competition. The competition is coordinated by Lecturer Dr. Jens Strackeljan at the Institute of Technical Mechanics at the Technical University of Clausthal.

The winners are Marcin Wojnarski at the University of Warsaw, Poland, first place; Dr. Bernhard Pfahringer at Waikato University in New Zeeland, second place; Dr. Dumitru-Iulian Nastac and Adrain Costea at the Computer Science Centre in Turku, Finland, third place. A total of twenty proposed solutions were received, some of them from the United States and Brazil, among other countries.

What was the problem posed for this competition? The data, 29 real input values as well as five real output values, each recorded at 15-minute intervals, were provided by Schott Glas, Mainz. These data were rescaled operational data for the control and indirect quality measurement of a glass melt over a period of fourteen weeks. The real physical meaning of these data was not revealed to the participants in the competition, since this, of course, is a company secret.

The measured values were supplied to the scientists in raw form; that is, they had not been pretreated or freed from interference, and noise was not suppressed. Thus, preparation of the data for modelling constituted a part of the given problem. A further difficulty was the frequent delay of hours or even days between the variation of an input signal and the “response” from the glass melting tank.

During the last two weeks, only the real input values to the system, that is, only the control targets of the process engineers and the sometimes unexpected, but measurable external effects, were known to the scientists. The manner in which the melt reacts to these inputs was not indicated to the scientists.

The problem for the scientists was to predict the behaviour of the glass melt. Their predictions and the derived mathematical correlations between the input and output values were subjected to an empirical test by comparison with the real values. The time-weighted deviation of the process behaviour predicted by the winners from the real values amounted to 0.3 per cent. They will present their results at the EUNITE Conference from 10th to 12th July in Oulu, Finland. As an expression of its gratitude, Schott Glas will award prizes in the amounts of 5000 € (1st place), 3000 € (2nd place), and 1000 € (3rd place).