A New Hybrid Fuzzy Time Series Forecasting Approach Based on Intelligent Optimization

In recent years, some intelligent techniques have been used in fuzzy time series approaches to improve the performance of these approaches. If intelligent techniques are utilized in fuzzification and defining fuzzy relations steps of fuzzy time series approaches, it makes these approaches systematic and it is not needed to make subjective decisions. Thus, the forecasting performance of fuzzy time series would increase. In fuzzification step, intelligent optimization techniques have been employed to partition universe of discourse into unequal intervals. Recently, artificial neural networks have widely been used in defining fuzzy relations step. When an intelligent optimization technique and a kind of artificial neural network are used in these two steps of a fuzzy time series, this fuzzy time series method has two optimization processes. One of them is an optimization process used to partition of universe of discourse. And the other one is training of artificial neural networks utilized to determine fuzzy relations. There are two separate objective functions in these two separate optimization processes. Therefore, the total error of the system is sum of errors produced by two different optimization techniques which are used to optimize two separate objective functions. A new hybrid high order fuzzy time series approach including only one optimization process is proposed in this study. In the proposed method, partition of universe of discourse and establishing fuzzy relations are performed at the same time by using particle swarm optimization algorithm. In order to define fuzzy relations, multiplicative neuron model is employed. Since the proposed approach includes only one optimization process with one objective function, error of the proposed approach is derived from only this optimization process. Therefore, it is expected that the forecasting performance of the proposed approach is high. As a result of an experimental study, it is shown that the proposed approach produces very accurate forecasts.