Physical-Statistical Hybrid Model for MIRROR PLANT

Abstract

MIRROR PLANT, the on-line plant simulator can perfectly simulate dynamic plant behavior. One of applications of MIRROR PLANT visualizes inside the plant where sensors cannot be actually installed. The plant operator can see the key performance index related with product quality in real time. And the operator can also see the prediction of the index. MIRROR PLANT uses physical plant modeling based on physical and chemical laws. However, only the physical model cannot express some phenomena in the plant, especially those which are not physically analyzed yet. We proposed a hybrid model of a physical model and a statistical model based on historical plant operation data. Combining the physical model and statistical model improves the estimation accuracy of the statistical model. In addition, by applying the just-in-time (JIT) modeling to the on-line update of the statistical model, it can deliver exact estimation even during rapid changes in the plant behavior.