Abstract

Synthetic climate time series are widely used in hydrologic modeling, analysis and design. There are several methods for generating time series and downscaling them to the required spatial and temporal scale. Nonetheless, many of these methods are unable to preserve local variability and spatial correlation, as well as other information of the General Circulation Model (GCM) output used to generate the time series. Moreover, these methods are very difficult to implement in semi-arid and highly variable climate regions. This work presents a weather generator method capable of generating future climate time series specially designed to address these drawbacks. The method considers three step: (1) GCMs probabilistic trend extraction, (2) locally coherent stochastic generation of annual precipitation and temperature data around the trend, and (3) annual data disaggregation into smaller time scales using the k-nearest neighbor (k-NN) method. The probabilistic trend allows quantifying explicitly the impact of the uncertainty in the generated trend and comparing it against the natural variability. The method was successfully applied to the Limarí, Maipo and Maule River basins, three river basins with very different climate conditions in the semiarid and Mediterranean region of Chile. Results show that the use of the k-NN resampling process allows preserving the strong seasonality typical of these areas, as the negative precipitations, commonly generated by alternative methods, are avoided. Despite the weather generator was designed for semi-arid regions, we show that it is valid to other climates.