AI-Driven Identification of Optimal Distribution Models for Extreme Precipitation Data in China

Choosing an appropriate distribution model to describe daily precipitation is vital. Previous studies focused mainly on overall accuracy; thus light-tailed distributions were frequently identified as suitable ones due to the adequacy of describing the bulk of data. Here, we leverage AI-driven techniques to identify the most appropriate distribution for modeling precipitation in China. We investigate five commonly adopted models using a dataset comprised of 618 station records. By utilizing a proposed statistic for measuring heaviness of observed tails’ behavior, we found that the demand for a heavy-tailed distribution decreases from the southeast to the northern interior region. Our results show that heavy-tailed models, i.e. the generalized gamma and Weibull, perform significantly better in capturing the upper tails of the distribution, which is critical for predicting severe infrastructure impacts. The relative frequencies of these two models being identified as optimal were 69.09% and 20.23%, respectively. While the light-tailed models severely underestimate extreme events. We further investigate two complex models derived from the initial five and recommend the exponential-Weibull mixture model due to its superior performance. This AI-driven approach not only enhances the accuracy of the extreme event modeling but also provides a robust framework for future climate risk assessments.