In this study, a fuzzy-based strategy for improvement of forecasting performance in data time series analysis is proposed. The designed methodology is target to seasonal autoregressive moving average processes modelling and can be applied to an wide range of real world applications. By means of hybrid approach based on a fuzzy version of correlation functions, the interpolating and the generalization capabilities of fuzzy systems are exploited in order to obtain a robust forecasting, even considering series with missing data points. In order to increase the algorithm accuracy, several design parameters were tested and optimized by computational tests. The following parameters are considered in this process: the length of the trajectory of the time series, the number of fuzzy sets, and the limit for activation of the support of the triangular fuzzy sets. It was observed that the membership function of triangular form lead to improved forecasting performance. A simulation to evaluate the accuracy of the forecasting of a fuzzy seasonal autoregressive model is described. To demonstrate the eectiveness of the proposed methodology, four case studies on data from some public data base was carried-out. The results conrm the improved performance of the proposed algorithm, allowing to obtain a reduced forecasting error in comparison to a conventional statistical methodology and fuzzy, for instance. The projections produced by the new method when subjected to fuzzy condence interval analysis showed satisfactory accuracy.
|
