Abstract: Using analysis data of NCEP FNL and forecast data of GRAPES_GFS as background fields, the error analysis of geopotential height(sounding height) data of Beijing sounding station are obtained from observation residuals, average deviations, standard deviation, probability density distributions, kurtosis coefficients, skewness coefficients, correlation coefficient and root mean square error. According to assessment results, data quality is marked, and parameters are solved according to results of the quality mark. Test results show whether based on NCEP FNL or GRAPES_GFS, the error of the sounding height is basically within ±5 dagpm, and the absolute value of observation residuals increases with the decrease of air pressure. Observation residuals below 100 hPa is basically within ±3 dagpm. Observation residuals are mostly negative at the top of 100 hPa. The average deviation, standard deviation, probability density distribution, kurtosis coefficient, skewness coefficient, correlation coefficient and root mean square error are analyzed and evaluated from characteristics and distribution characteristics of the seasonal error, all of which show that the quality of data at height of detection potential is good, and each parameter is close to their optimal state. However, at the high level (10-30 hPa), the average deviation and standard deviation show obviously that the evaluation result of GRAPES_GFS is better than that of NCEP FNL, and the other parameters are basically the same and the difference is small. The average deviations plus standard deviation of two times is selected as the suspicious threshold value of the potential height at a single moment, and average deviation plus standard deviation is selected as the error threshold of the potential height. This choice is not only meaningful in mathematical statistics, but also shows that the threshold value is based on the background field error feature and self-adaptive threshold value, which can help to find out the true error point for correction.