Abstract: Quality control experiments about the surface observation data assimilation in Sichuan Province were carried out by using of the GRAPES assimilation and prediction system.0-48 hprecipitation forecasts of three heavy rainfall process that occurred in Sichuan Province at 3 September 2004, 2July 2005, 20July 2008 were analyzed.The main conclusions were as follows:Firstly, the analysis increment field were changed with the error multiple of surface assimilation variable;when the error multiple of variable H was taken 4 times, 80% data could be admitted to the assimilation system;when the error multiple of variable U、V、Q were taken 2 times, 90% data could be admitted;Secondly, from the comparative analysis results of the object function and its gradient decreased with iterative steps, they showed that when the error multiple of surface assimilation variable were taken 4 times and its under, their object functions could meet the convergence condition, then obtained the analysis field;Thirdly, the results of rainfall prediction would be worse with the error multiple of variable H increased and more H information of surface observation data assimilating;the results of rainfall prediction would be better with more U、V、Q information of surface observed data assimilating, and the prediction of rainfall band are the same with their error multiple were taken 3 times and its upper;Finally, under the condition of assimilating more surface observed data and better assimilation quality, the error multiple of variable H、U、V、Q are 4、3、3、3 times.