A Novel Fuzzy Time Series Forecasting Model by Similarity Measurement and Fuzzy Inference

The definition and forecasting framework of fuzzy time series are first proposed by Song and Chissom in 1993. The forecasting framework includes five steps: (1) determine and partition the universe of discourse, (2) define fuzzy sets and fuzzify the time series, (3) derive fuzzy relationships, (4) forecast, and (5) defuzzify the forecasting outputs. Many researchers using the fuzzy logical rule to construct the fuzzy relationship on step (3) and using 'table-look-up' method on step (4). However, there are three shortcomings in step (3) and (4), which are the rule redundancy, the rule length redundancy, and ignored the characters of fuzzy sets. In this paper, we propose similarity measurement method to prune the length of rule and apply fuzzy inference to raise the character of fuzzy sets on forecast step. And more, the quantile partition is applied on partition the universe of discourse. Final, the proposed model achieves the best accuracy in the forecasting average temperature of the daily average temperature in Slovakia of Eastern Europe.