A stochastic weather generator applied to hydrological models in climate impact analysis

A pattern recognition methodology for estimating local climate variables such as regional precipitation and air temperature using local observation and scenario information provided by GCMs is presented. We have adopted a three step approach: (a) Feature information extraction of climate variables, where weather patterns are expanded by the Karhunen-Loeve (K-L) orthogonal functional series; (b) Grey associative clustering of the feature vectors; (3) Stochastic weather generation by a Monte Carlo simulation. The methods described in this paper were verified using the temperature and precipitation data set of Wuhan, Yangtze river basin and the Shun Tian catchment, Dongjiang River in China. The proposed method yields good stochastic simulations and also provides useful information on temporal or spatial downscaling and uncertainty.