| Summary: | 碩士 === 淡江大學 === 水資源及環境工程學系 === 88 === In the past, differenced and detrended method are usually used in analyzing long-memory hydrological time series. Recently, a k-factor GARMA model was proposed to model long-memory time series, and its forecasting results are better than conventional methods. The major purpose of this study is to examine the suitability of this model, and then apply it to monthly riverflow data of Taiwan.
In this study, approximate maximum likelihood method proposed by Hosking(1984) is used to estimate the parameters of k-factor GARMA model, and AIC and BIC are used for model identification. Besides, MSE﹑MAPE and UI are used as criteria for comparison of forecasting. The accuracy of approximate maximum likelihood method is affected by sample size. The larger the sample size, the better parameter estimation accuracy. The results of model identification are affected by parameter and sample size. The BIC criterion has better results in model identification than AIC criterion. On the other hand, the performance of k-factor GARMA model is better than detrended model in preserving skewness when real data are analyzed. To the opposite, detrended model fits better to the mean and variance than k-factor GARMA model. The forecasting results show that k-factor GARMA model is better than detrended model, especially when streamflow is low.
In conclusion, k-factor GARMA model has better forecasting ability than detrended model, and improve preserving skewness. In general, the k-factor GARMA model is a reasonable model for the monthly riverflow data of Taiwan.
|
|---|
