Model, Time Series Forecasting Based On Fuzzy Theory

Fuzzy time series have received more attention due to their abilities to deal with the vagueness and incompleteness inherent in time series data, and applied to predict enrollment, temperature, stock prices and so on. In the last decade, various improvements have been proposed, and most of them are improvements based on the model of Song in 1994 proposed, such as high-order model and heuristic model. They were all aim at improving the forecast accuracy, or reducing the computational complexity.However, for the improvement based on proposed heuristic model, and the uncertainty inherent in fuzzy time series and long-term forecasting are still less. Therefore, this paper studies on the following issues:Firstly, how to based on existing heuristic model, expansion of heuristic rules to make the prediction performance better; Secondly, how to deal with the uncertainty state transitions exist in fuzzy time-series data, offer beneficial reference for the performance improvement; Thirdly, according to most existing time series models are short-term forecasting model, proposed the long-term forecast model of fuzzy time series.Therefore, this paper done as follows based on these aspects of work:(1) Based on existing first-order model and higher-order model, as well as heuristic model, proposed an improved heuristic time-invariant prediction model. This uses prediction accuracy of model observations to train the trend predictor in the training phase, and uses this trend predictor to generate forecasting values in test phase, so as to enhance the prediction performance.(2) For the uncertainty inherent in the state transition fuzzy rules and the rules construction problem, proposed a deterministic model. Based on the analysis of the existing models, by adding back-tracking scheme to build a deterministic transition rule, so as to make the prediction accuracy improved.(3) For overcoming the shortcoming that the existing models are mostly limited to short-term forecasting, proposed a long-term forecasting model based on vector quantization. By adding sliding windows scheme and vector quantization technique, to support forecasting if there are no matching historical patterns, which is usually the case of long-term forecasting.