On One Approach to Predictive Modeling Based on Monitoring Data

Abstract

The paper proposes one approach to predict the behavior of complex dynamic systems based on monitoring data presented as time series. The general idea of the approach is a decomposition of a prediction task of the behavior of a complex dynamic system possessing substantial structural and parametric uncertainty. The prediction task is performed in two stages. The first stage realizes the retrospective situational modeling for individual time periods resulting in a set of situational models in the form of simple regressions. The second stage comprises inductive modeling based on the previous situational modeling results. According to the approach, the prediction by means of such combined situational-inductive modeling is aimed at establishing relevant situational models of regression type being adequate in the future within certain limited time periods. The approach allows using simultaneously both the principle of optimization in modeling and the principle of adaptation to situational changes occurring in dynamic systems.