Stochastic generation of daily precipitation amounts: review and evaluation of different models

The present study first reviews the performance of different models in generating daily precipitation amounts. Eight models with different levels of complexity are then selected to simulate daily precipitation for 35 stations across the world. All 8 models adequately reproduce the observed mean precipitation at daily, monthly and annual scales, while all of them underestimate the standard deviation of monthly and annual precipitation. However, the compound distributions are generally better than the single distributions at reducing the variance overdispersion, with the exception of the skewed normal (SN) distribution. The nonparametric kernel density estimation (KDE) is consistently better than all the parametric distributions. With the exception of the SN distribution, all the single distributions underestimate the upper tail of daily precipitation distribution. However, the generalized Pareto distribution- based compound distributions provide a reasonable performance for simulating the upper tail, even though they are slightly worse than the KDE, which displays the best performance. Overall, the compound distributions generally perform better than the single distributions, and the nonparametric KDE performs better than the parametric distributions. However, the complicated structure of the compound distribution and of the KDE and the limited extrapolation ability of the KDE may restrict their application to climate change impact studies. The 3-parameter SN distribution displays a similar or even slightly better performance than the compound distributions, and this distribution may be the first choice to be incorporated into a weather generator for studying climate change impacts, especially for risk-related assessments.