CR 14:75-87 (2000)  -  doi:10.3354/cr014075

Improvements of stochastic weather data generators for diverse climates

Henry N. Hayhoe*

Eastern Cereal and Oilseed Research Centre, Research Branch, Agriculture and Agri-Food Canada, Ottawa, Ontario K1A 0C6, Canada

ABSTRACT: Weather data generators can be used to produce long periods of synthetic weather records from a limited amount of input data. They provide a useful tool for supplementing existing climate data for risk assessment and decision support systems. They also can be adapted for use in climate change studies where only limited data are available on a course spatial grid. A number of studies have demonstrated the failure of data produced with stochastic weather data generators to mimic the statistical properties of observed daily climate data. Differences may be large enough to limit the usefulness of the output data. This study suggests that the lack-of-fit is related to simplifying assumptions incorporated in these models which may be acceptable in some climates but lead to significant discrepancies when applied under very different climate conditions. Methods are proposed to address these limitations. The challenge related to choosing an appropriate probability distribution for each climate variable is addressed by approximating the probability distribution of the data using an observed frequency distribution. It is becoming increasingly clear that the correlation between climate variables may differ significantly on a seasonal and regional basis. To account for this, the interrelationship between climate parameters is evaluated for each site once every 2 mo. The way the relationship between the wet or dry status of the day and temperature and solar radiation is accounted for is improved in the model. Observed climate data for 3 selected Canadian climatological stations are used to assess the proposed weather data generator. The results indicate an overall improvement in the correspondence between the means, standard deviations, correlations, probability distributions and extreme values of the observed and estimated weather data series when compared with previously reported results. Related parameters such as the length of wet periods, the length of dry periods and the length of frost-free periods are also found to compare favourably with the observed values.

KEY WORDS: Markov chain · Stochastic process · Probability distribution · Correlation · Climate change

Full text in pdf format