| Summary: | 博士 === 國立成功大學 === 工業與資訊管理學系碩博士班 === 98 === With the fast growth of the temporal data evolving over time, a great many varieties of time series forecasting techniques have been studied for providing precise results efficiently, which include traditional statistics, artificial neural networks and fuzzy methods. However, the forecasting capabilities of most of the aforementioned techniques are limited to short-term time spans in which a single future value is forecasted in one step. Nevertheless, there is an increasing need for long-term forecasting going many time steps in advance. For instance, when forecasting the monthly energy consumption of households, it would be more useful to predict the long-term trend and all monthly values for one season in a single step. Long-term forecasting is a task that is difficult to achieve because information is unavailable for the unknown future time steps. Although there have been long-term forecasting models proposed in the recent years, they mainly deal with numeric historical data. However, in practice uncertainties, such as incompleteness, impreciseness and ambiguousness are likely to be widespread in real-world data, and hinder forecasting accuracy, thus limiting the applicability of these models. To deal with such uncertainties, in this study we propose a fuzzy pattern linguistics based long-term time series forecasting model, which is designed for vector long term time series forecasting by combining the concepts of both fuzzy set theory and time series forecasting technique.
In the proposed method, we incorporate the well-recognized sliding-window scheme to extract features of interest in a time series. Fuzzy c-means (FCM) clustering is then used to handle interval partitioning, in which we take the density and uncertainty of data points into account, and unequal-size intervals are constructed. In this work, the temporal relationships in a fuzzy time series are deterministically extracted and represented as certain transition rules to facilitate the forecasting. In the case when no historical certain rules are available for the unseen time series, the forecasting is enhanced by applying vector quantization (VQ). Finally, due to the uncertainties of the initially assigned membership degrees in FCM clustering, Monte Carlo simulation is utilized to verify the reliability of the proposed model. To validate the effectiveness of the proposed method, we conduct real-world experiments on monthly accumulated rainfall and daily temperature in Taiwan. The results indicate that the proposed method provides better forecasting ability than traditional statistics (ARIMA), artificial neural networks (BPN) and fuzzy methods (Deterministic Fuzzy Time Series).
|
|---|
