180-1 Stochastic Weather Generators: Applications, Modeling Approaches, and Model Evaluation.

See more from this Division: A11 Biometry
See more from this Session: Symposium--Time Series Analysis and Forecasting in Agriculture Research
Tuesday, November 2, 2010: 1:00 PM
Long Beach Convention Center, Room 102B, First Floor
Share |

Justin T. Schoof, Southern Illinois University, Carbondale, IL
Stochastic weather generators (WGs) are statistical models that produce artificial sequences of meteorological variables.  In the last three decades, WGs have emerged as a critical tool for crop modeling applications, but have also been identified as a means of producing daily sequences consistent with climate change scenarios produced by coupled climate models.  This presentation will provide an overview of common WG applications with a focus on methodological approaches and resulting strengths and weakness.  Most WG applications use variations of the original WGEN model introduced by Richardson and Wright (1984).  WGEN produces daily sequences of precipitation (occurrence and intensity), temperature (maximum and minimum) and total solar radiation.  Precipitation occurrence and non-zero intensity are governed by a 1st order Markov chain and Gamma distribution, respectively.  The remaining variables are generated using a first-order autoregressive process, with dependence on wet/dry status.  Common critiques of the WGEN-type model are that (1) the lengths of wet and dry spells are not consistent with those observed in some regions, (2) large precipitation events are underestimated, and (3) the Gaussian distribution is inappropriate for the non-precipitation variables, which are generally skewed.  These findings have led to several developments aimed at improving the characteristics of the generated series.  For precipitation occurrence, higher order Markov chains have been advocated and alternative methods, such as spell length approaches, have been developed.  Alternative distributions for non-zero precipitation amounts have also been proposed to improve generation of precipitation extremes.  Similarly, new models for non-precipitation variables have been proposed to address the non-normal distributions of temperature and solar radiation.  Methodological approaches have included the use of semi-empirical distributions, resampling strategies and spectral techniques.  Lastly, several recent applications have included non-traditional variables such as humidity and wind.  The most important characteristics of the generated sequence are likely to vary from one application to the next.  This overview will provide specific examples of the types of improvements achieved through use of different methodological approaches.
See more from this Division: A11 Biometry
See more from this Session: Symposium--Time Series Analysis and Forecasting in Agriculture Research