Abstract
Stochastic weather generators are designed to produce synthetic sequences that are commonly used for risk discovery, as they would contain rare events that can lead to potentially catastrophic impacts on the environment, or even human lives. These time series are sometimes used as inputs to rainfall-runoff models to simulate the hydrological impacts of these rare events. This paper puts forward a method that evaluates the usefulness of weather generators by assessing how the statistical properties of simulated precipitation, temperatures, and streamflow deviate from those of observations. This is achieved by plotting a large ensemble of (1) synthetic precipitation and temperature time series in a Climate Statistics Space, and (2) hydrological indices using simulated streamflow data in a Risk and Performance Indicators Space. Assessment of weather generator's performance is based on visual inspection and the Mahalanobis distance between statistics derived from observations and simulations. A case study was carried out on the South Nations watershed in Ontario, Canada, using five different weather generators: two versions of a single-site Weather Generator, two versions of a multi-site Weather Generator (MulGETS) and the K-Nearest Neighbour weather generator (k-nn). Results show that the MulGETS model often outperformed the other weather generators for that particular study area because: (a) the observations were well centered within apoint cloud of the synthetically-generated time series in both spaces, and (b) the points generated using MulGETS had a smaller Mahalanobis distance to the observations than those generated with the other weather generators. The k-nn weather generator performed particularly well in simulating temperature variables, but was poor at modelling precipitation and streamflow statistics.