| Summary: | 碩士 === 國立臺灣科技大學 === 電機工程系 === 90 === In recent decades, predicting future becomes an important issue. Traditional forecasting approaches are based on all the available training data including the nearest or far away from the present. Those approaches are referred to as the global prediction scheme. On the other hand, those prediction approaches that only construct their prediction model based on the most recent data are referred to as the local prediction schemes. In most cases, local prediction schemes can have better prediction performances than that of global prediction schemes for time series. In the literature, an approach, Markov Fourier Grey model (MFGM) has been proposed to incorporate global information based on local prediction schemes. In traditional forecasting, people may want to predict the next data and this kind of prediction is called one-step prediction. Nevertheless, we may also need to make multi-step prediction. From our simulation, it can be found that those local prediction schemes or MFGM even can have nice performance in one-step prediction, they usually have awful performance for multi-step prediction. In this study, we intend to study various approaches in combining local and global prediction approaches. They are traditional Kalman filter approaches, network based filter fusion approaches, and network based residual modeling approaches. Neural networks are widely used to predict time series. In our study, neural networks and SONFIN are employed as global prediction schemes and Fourier Grey Model (FGM) is employed as local prediction schemes. There are three example time series studied here. The first example of time series is a nonlinear function. The second time series is a history of the stock market. The third example is the Mackey-glass chaotic time series. The performances of using those proposed approaches for predicting those three time series are reported and analyzed in this thesis.
|
|---|
