Impact of number of realizations on the suitability of simulated weather data for hydrologic and environmental applications

Authors: GUO, TIAN - Purdue University; MEHAN, SUSHANT - Purdue University; GITAU, MARGARET - Purdue University; WANG, QI - Purdue University; KUCZEK, THOMAS - Purdue University; Flanagan, Dennis

Submitted to: Stochastic Environmental Research and Risk Assessment (SERRA)
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 11/27/2017
Publication Date: 12/6/2017
Citation: Guo, T., Mehan, S., Gitau, M.W., Wang, Q., Kuczek, T., Flanagan, D.C. 2017. Impact of number of realizations on the suitability of simulated weather data for hydrologic and environmental applications. Stochastic Environmental Research and Risk Assessment (SERRA). 32(8):2405-2421. https://doi.org/10.1007/s00477-017-1498-5.
DOI: https://doi.org/10.1007/s00477-017-1498-5

Interpretive Summary: Weather is an important driver of many processes at the soil surface, in particular infiltration, runoff, soil loss, and agricultural chemical losses. Monitoring of runoff and pollutant losses from agricultural production areas is difficult, expensive, and impractical for large areas, so usually computer simulation models are utilized to determine potential losses of runoff water, sediments, nutrients, and pesticides in response to daily weather drivers (precipitation, air temperature, etc.). In order to provide input to these natural resource models, other computer programs known as “climate generators” are used, that are meant to produce daily values for rainfall and temperatures that match those from statistical records of observed data from weather stations. In this research study, 3 different climate generator programs were evaluated using data from a weather station in Fort Wayne, Indiana with 50 years of observed data. We found that for all 3 generators, creating a single prediction of a string of daily climate was not sufficient to produce statistical values that would match those of the observed data at this station, while creation of 25 or more strings was sufficient. These results are important and impact scientists, university faculty, students, conservation agency personnel and others that use results from climate generators in application of natural resource models. This indicates that when using these types of climate generators and models, it may be necessary to conduct multiple model applications with more than a single realization of predicted daily weather. Minimally, users need to be aware of the implications of use of a single realization from a climate generator.

Technical Abstract: Stochastic weather generators are widely used in hydrological, environmental, and agricultural applications to simulate and forecast weather time series. However, such stochastic processes usually produce random outputs hence the question on how representative the generated data are if obtained from only one simulation run (realization) as is common practice. In this study, the impact of different numbers of realizations (1, 10, 25, 50, and 100) on the suitability of generated weather data was investigated. Specifically, 50 years of daily precipitation, and maximum and minimum temperatures were generated for the Fort Wayne weather station in Indiana, using three widely used weather generators, CLIGEN, LARSWG and WeaGETS. Generated results were compared with 50 years of observed data. For all three generators, the analyses showed that one realization of data for 50 years of daily precipitation, and maximum and minimum temperatures may not be representative enough to capture essential statistical characteristics of the climate. Results from the three generators captured the essential statistical characteristics of the climate when the number of realizations was increased from 1 to 25, 50 or 100. Performance did not improve substantially when realizations were increased above 25. However, an assessment based on 10 realizations showed that the generated results may not be able to capture some essential statistical characteristics, which were captured by 25 realizations. Results suggest the need for more than a single realization when generating weather data and subsequently utilizing in other models, to obtain suitable representations of weather for a location.