Research On The Fuzzy Time Series Model Based On Particle Swarm Optimization With Variable Parameters

In view of the time series with uncertainty or vagueness, the traditional forecasting methods has poor adaptability, so many scholars focus on the fuzzy time series prediction problems. At present, the fuzzy time series has been used in the number of visitors forecast, temperature forecasts, stock market forecasts, the number of Internet users forecast and many other areas. In order to improve prediction accuracy of fuzzy time series models, scholars optimized the domain partition and fuzzy rules that were proposed in the four-step model by Song and Chrisom[1-3] in 1993.Interval division is the cornerstone of fuzzy time series models, in the earlier studies interval division were mostly equally spaced domain partition, but the forecast accuracy of split equally spaced domain method is low, with strong deliberate, so many scholars have suggested the non-equal interval classification field and achieved a better predictive accuracy. As a number of advanced intelligent optimization algorithm (genetic algorithms, Ant Colony Optimization, Particle Swarm Optimization) are applied to the prediction model, effectively optimize the algorithm into the prediction model to improve the forecasting model research. This article use the Particle Swarm Optimization algorithm with variable parameters to divide the domain.Using the segmentation by PSO algorithm got as the endpoints of the domain partition, on this basis, defining the fuzzy sets and fuzzifying the time series on the universe of discourse, and constructing fuzzy logic relationships. Then, the characteristic coefficients of the minor premises about inference rules are calculated by the characteristic expansion method. Fuzzy rules are reducted and optimized. Finally, the prediction results are obtained by using the defuzzify.By adjusting parameters, we combined the PSO with the constriction factor algorithm with characteristic expansion method applied to Number of enrollment of the University of Alabama between 1971 and 1992 and Amazon’s stock closed in first half of 2013, it verify the effectiveness of the algorithm proposed in this article.