Develop a multi-linear-trend fuzzy information granule based short-term time series forecasting model with k-medoids clustering

Abstract

In fuzzy information granule (FIG) based short-term forecasting models, the constructed FIG focuses on one of two tasks: capture data characteristic and improve semantic description at a common time concept (time interpretability). For accomplishing both tasks at once, a multilinear-trend fuzzy information granulation algorithm is raised with three-stage time span adjustments, which consistent with the multi-linear-trend information of data and time interpretability. The multilinear-trend FIGs granulated through this algorithm are composed of different number of unequal-size linear FIGs. In order to identify the difference of such granules, k-medoids based multilinear-trend FIG clustering algorithm is put forward, and a cluster label series corresponds to a series of multilinear-trend FIGs can be obtained, where each cluster label represents one kind of multi-linear-trend patterns. Following these, the correlation among patterns will be depicted by a new type fuzzy association rule (FAR), cluster label association rule, which characterizedby simple form so that suitable for practical application. Combing the proposed algorithms with FAR, a novel model for short-term forecasting at granular level is designed, it makes results from available FARs and reduces cumulative errors in prediction. The better forecasting performance of this model has been verified after comparing it with existing models in experiments.