Abstract: Traditional GM(1,1) and time series analysis method were integrated together to establish the groundwater level forecast model.  The fading memory least square methods was applied in the process of confirming the parameters of GM(1,1) in order to more factually reflect the phenomena of the groundwater level's statistic rule varying with the time. The periods components were analyzed with the variance analysis method and the starting phases were optimized. Under the condition that more than one F-statistic values of periods were over the critical values of the selected confidence degree, what rules can be adopted to decide which period should be first selected. The rule was put forward that the ratio of F-statistic value of  selected period's being divided by the critical value of the examination interval should be the greatest. Finally, the yearly highest groundwater levels in Longyan Basin in the west of  Fujian Province were forecast with this method.

Key words: groundwater level forecast, GM(1, 1), variance analysis, starting phase optimization, time series analysis