III. Model mixing

Proper prediction of system output provides valuable information for control or process supervision. For model based techniques the knowledge about the model structure is essential.
There exist processes, which are expected to behave according to several physical laws at once and each of these laws can imply a model which should match the measured process data. Nevertheless predictions calculated from these single models can be unusable because of imperfect measurement, unmeasurable disturbances or even because of principal divergence of the process behavior from the supposed model. Mixing model technique results in the weighted prediction which can be more reliable than prediction based on particular models. The research is motivated by existing problems from the field of metal rolling. Two approaches are being tested:

• Separated explicit models of the process are used. Their parameters are identified recursively and particular predictions are used as inputs for the superordinate model. Resulting prediction is weighted by its parameters.
• The system is modelled by a probabilistic mixture the components of which represent single models. Model parameters and weights are estimated at once.