13th Conference on Applied Climatology and the 10th Conference on Aviation, Range, and Aerospace Meteorology

Tuesday, 14 May 2002: 9:30 AM
Implementing quality control techniques for random number generators to improve stochastic weather generators: the CLIGEN experience.
Charles R. Meyer, USDA/ARS, West Lafayette, IN; and C. Renschler and R. C. Vining
Poster PDF (294.7 kB)
As with most other stochastic weather generators, CLIGEN simulates distributions of daily weather parameters based on monthly statistical values observed at a particular site. The quality of the results produced by stochastic models like CLIGEN depends directly upon the quality of the distributions produced by their random number generators. Statistical tests with various random number generators and CLIGEN version 4.2 demonstrated serious flaws in these distributions. Lack of quality assurance for these random numbers has potentially serious implications for these programs and the performance of simulation models that depend on them. Quality control software was written into CLIGEN version 5.x to compute the probability that the mean and standard deviation of the random numbers driving CLIGEN were standard normal as assumed, and reject those that were not. An arbitrary 50 percent probability threshold was employed. We compared the historical monthly means of daily values from four climatically diverse sites in the United States (Indianapolis, IN, College Station, TX, Moscow, ID, Tucson, AZ). CLIGEN version 5.x employing quality control techniques improved reproducibility of climatic parameters compared to CLIGEN version 4.2. We recommend this approach be used with other random number generators in comparable situations with stochastic climate generators.

Supplementary URL: http://horizon.nserl.purdue.edu/Cligen/