Based on the analysis of daily p recip itation of 62 weather stations in Gansu Province from 1960 to 2005, it educed the characteristics of spatial variation of different intensity p recip itation days and p recip itation relative extremes, and those can reflect the structure of p recip itaiton. The results show that there isn’t any homogeneous variation over the whole p rovince for each indicator, and the spatial distinction of p recip itation variation is great. Thus, the regional average of p recip itation ismeaningless and the structure ismore powerful to exp lain the change of p recip itation, for which decomposed the dim unitary situation into vivid particular phenomena. The farther research on spatial variation of p recip itation structure indicates that a significant variation of p recip itation in a specific region is a synthetical consequence of the interactions among all indicators. This synthesis p rocedure is very comp licated, and usually has not broad rep resentation. Even nonsignificant variation of each indicator can bring forth a significant variation of p recip itation.

The predictionmodel for the highestand lowest temperature is established by the CMSVM regressionmethod based on ECMWF data and the extreme temperature observations from 8 automaticweather stations inQingyang ofGansu Province from 2003 to2007.The effect test in 2008 operation forecast indicated that the rate ofaverage forecastaccuracy is64% and 71% for the highestand lowest temperature in 5 days, which has the good instruction for the real-timeweather service. The forecast accuracy rate of the highest and lowest temperature decreased with the forecast time increase, and the accuracy rate for the lowest temperature is higher than thatof the highest temperature. The forecast effect for the lowest temperature is better in spring and summer than that in autume and winter, but it is opposite for the highest temperature. The absolute and the average forecasterrors for the highestand lowest temperature are added with the forecast time increase, and the absolute forecasterror for the highest temperature is bigger than thatof the lowestbut the average error is close and positive. The forecast effect test in Septemper of 2008 indicated that the forecast values presented the similar trend to real time observations for the highest and lowest temperature.