Research On One Class Of Hybrid Forecasting Models For Fuzzy Time Series

In recent years, fuzzy time series forecasting methods have been widely applied in many domains, and the studies on fuzzy time series models are getting more attention. Related studies mainly focus on three factors, namely, the partition of discourse, the content of forecasting rules, and the methods of defuzzification, all of which greatly influence the prediction accuracy. In order to improve the prediction accuracy, two novel hybrid forecasting models are proposed based on the existing models in this paper, i.e. Adaptive Forecasting models with Classical-fuzzy-time-series and Heuristic methods, and a new hybrid forecasting model with Fuzzy c-means clustering and Genetic algorithm based on the Adaptive Forecasting models with Classical-fuzzy-time-series and Heuristic methods.Firstly, an Adaptive Forecasting model with Classical-fuzzy-time-series and Heuristic methods is proposed in this paper. The proposed model automatically adapts the analysis window size of fuzzy time series based on the prediction accuracy in the training phase, overcoming the problem of some models used fixed analysis window for forecasting to improve the prediction accuracy. In addition, we tend to only consider the current state and ignore the next state in the prediction process, which could be an important factor affected the prediction accuracy. To deal with the impact of this factor, in the testing phase of the proposed model, it firstly use Mamdani reasoning method of the Classical-fuzzy-time-series forecasting methods to predict the fuzzification result of the next state, then an improved Heuristic algorithm is employed to calculate the forecasting values.Secondly, a new hybrid forecasting model with Fuzzy c-means clustering and Genetic algorithm is proposed based on Adaptive Forecasting models with Classical-fuzzy-time-series and Heuristic methods in this paper. The proposed model firstly use Fuzzy c-means clustering fuzzification the history data and classify them, then adopts the Genetic algorithm to optimize certain interval length in the fuzzification phase, dealing with the problem of existing information loss and explanation power decreasing with dividing the universe of discourse into equal intervals. Then, in the forecasting phase of the model, calculating the forecasting values with the forecasting method of the Adaptive Forecasting models with Classical-fuzzy-time-series and Heuristic methods based on the obtained fuzzification results and optimal intervals. The two proposed models are employed in different time series in the simulation experiment, including the Enrollment of the University of Alabama, TAIFEX(Taiwan Futures Exchange) data, Mackey-Glass time series, et al. Comparing the forecasting results with other forecasting models in the literature to verify the models’ performance in this paper, the experiment results show that, the proposed models have higher forecasting accuracy and achieve better prediction performance.