The statistics of 6-hour forecast errors for z, u, and v derived from the global data assimilation system at the Central Weather Bureau in Taiwan are presented. One point moments, including mean, standard deviation, skewness, and kurtosisi of the forecast errors, are calculated at radiosonde stations to evaluate the statistical properties and define how close the distribution of the forecast error is to the Gaussian distribution. The degree to which the analyses fit the observation is also examined.
The overall evaluations with respect to different domains show that the lower order statistics, mean and standard deviation, are reasonable and comparable to the results of other operational centers. The higher order statistics show that the distributions of the forecast error form an approximate Gaussian distribution.
The spatial distribution of the one point moment shows that the mean and standard deviation of forecast errors are sensitive to the orographic effect (e.g., the Tibetan Plateau), the Asia and North American monsoon activities, and the mid-latitude disturbances. The pattern of the mean and standard deviation exhibits large-scale variability, which may be attributed to the background errors and suggest that the model error is dominated by large scales. The skewness and kurtosis have many local extremes, suggesting that observational errors dominate these higher order statistics.