E2GK-pro: An Evidential Evolving Multimodeling Approach for Systems Behavior Prediction

Abstract

Nonlinear dynamic systems identification and nonlinear dynamic behavior prediction are important tasks in several areas of industrial applications. Multiple works proposed multimodel-based approaches to model nonlinear systems. Multimodeling permits to blend different model types together to form hybrid models. It advocates the use of existing, well known model types within the same model structure. Recently, a multi modeling strategy based on belief functions theory was developed based on a fuzzy rule based system. We propose a different approach of this latter taking advantage of new efficient evidential clustering algorithms for the determination of the local models and the assessment of the global model. In particular, we propose an online procedure based on the Evidential Evolv- ing Gustafsson-Kessel (E2GK) algorithm that ensures an evolving partitioning of the data into clusters that correspond to operating regions of the global system. Thus the estimation of the local models is dynamically performed by upgrading and modifying their parameters while the data arrive. Each local model is weighted by a belief mass provided by E2GK, and the global model (multimodel) is a combination of all the local models.