Fuzzy Granulation Based Forecasting of Time Series

In this paper, a novel fuzzy granulation based forecasting method is presented for time series. This method includes two steps: granular modeling and forecasting. In granular modeling step, the given time series is first partitioned in terms of data condense degree into segments (windows) with different widths, then after optimally constructing fuzzy granule on each window, a fuzzy granular time series best fitting the original time series is obtained. In the forecasting step, we first fix the linguistic depiction of each granule and build the forecasting rules by mining the fuzzy relationship between the adjacent granules in the granular time series obtained in the first step. After that, we finish the forecasting by means of the forecasting rules. This fuzzy granulation based method can give not only linguistic prediction but also crisp prediction. The main difference of this method from the existing methods is that it realizes the granulation by optimization where the granules correspond to different widths. Thus the model presented here can be regarded as a universal one. Experiment carried on the enrollment data of Alabama University illustrates the good performance of the new method.