A Multi-Site Stochastic Weather Generator for High-Frequency Precipitation Using Censored Skew-Symmetric Distribution

TL;DR

This paper introduces a novel multi-site stochastic weather generator tailored for high-frequency precipitation data, effectively capturing skewness, heavy tails, and zero-inflation, with applications demonstrated on dense gauge network data from Lausanne, Switzerland.

Contribution

It proposes a censored non-Gaussian vector autoregression model with skew-symmetric dynamics for high-frequency precipitation generation, improving modeling of complex precipitation patterns.

Findings

Accurately reproduces high-frequency precipitation data.

Provides reliable predictions with interpretable uncertainties.

Outperforms traditional models in capturing skewness and heavy tails.

Abstract

Stochastic weather generators (SWGs) are digital twins of complex weather processes and widely used in agriculture and urban design. Due to improved measuring instruments, an accurate SWG for high-frequency precipitation is now possible. However, high-frequency precipitation data are more zero-inflated, skewed, and heavy-tailed than common (hourly or daily) precipitation data. Therefore, classical methods that either model precipitation occurrence independently of their intensity or assume that the precipitation follows a censored meta-Gaussian process may not be appropriate. In this work, we propose a novel multi-site precipitation generator that drives both occurrence and intensity by a censored non-Gaussian vector autoregression model with skew-symmetric dynamics. The proposed SWG is advantageous in modeling skewed and heavy-tailed data with direct physical and statistical interpretations. We apply the proposed model to 30-second precipitation based on the data obtained from a dense gauge network in Lausanne, Switzerland. In addition to reproducing the high-frequency precipitation, the model can provide accurate predictions as the long short-term memory (LSTM) network but with uncertainties and more interpretable results.