Key Words:
Autocorrelation Function, ARIMA Model, Land Subsidence, Stochastic Process, Stationality

Abstract:
This study investigates the application of time series analysis methods for forecasting groundwater levels. The study site is located in western Taiwan where serious land subsidence has occurred. A series of monthly groundwater level observations made during the period 1993 and 1999 is used for the experiments. Univariate time series models, including ARIMA models and the time series decomposition method, are applied and the resulting accuracy is compared. Empirical results indicate that groundwater level data series in this study are cyclical. ARIMA models generate more accurate forecasts. The forecasting of ARIMA models presents the characteristics of trend and seasonal variation.

1. Introduction
Groundwater level models provide useful information for land subsidence forecasts. The Univariate Box-Jenkins (UBJ) ARIMA analysis (Box, Jenkins and Reinsel, 1994) has been used in many applications, such as medical, environmental, financial, and engineering applications (Abdel-Aal and Mangoud, 1998; Kumar and Jain, 1999; Mitosek, 2000). A comparative study between ARIMA models and the time series decomposition model for forecasting groundwater levels is discussed in this article.

2. Site Description and Data Collection
Records of groundwater levels were compiled for wells in a monitoring network in the Cho-Shui alluvial fan near southwest Taiwan, starting around 1993. Groundwater-levels were generally measured once a month.

3. Arima Modeling of the Groundwater Level
The statistical package MINITAB for Windows is employed to develop the ARIMA models. There are four stages in the modeling process (Bowerman and O'Connell, 1993), i.e. identification, estimation, diagnostic checking, and forecasting.