Coupled annual and daily multivariate and multisite stochastic weather generator to preserve low- and high-frequency variability to assess climate vulnerability

Abstract

Multivariate and multisite stochastic weather generators have been proposed to produce an ensemble of climate time series but are often limited to preserving low-frequency variability of climate variables as well as producing extreme rainfall events. This study presents a new, two-stage, multivariate, multisite weather generator by coupling annual and daily weather generators. A daily weather generator is first developed using heavy tailed distribution, spatial tail dependence, and multivariate autoregressive models. For the second stage, annual climate variables over the region are modeled using a wavelet decomposition approach coupled with a multivariate autoregressive model. The generated annual time series are used to reconstruct daily simulations to embed multiple low-frequency oscillations in a daily time series. The proposed weather generator is applied to the Geum River Basin in South Korea, and its performance is compared to that of a nested simplified model. Results show that the proposed model performs well with respect to reproducing marginal distributional attributes, multisite dependencies, and climate variability at the daily and annual scales. Lastly, the weather generator adopts a quantile mapping procedure to incorporate long-term distributional changes into generated climate sequences for use in climate change assessments. Results show that inter-annual variability is also well preserved while climate sequences are adjusted by various alterations.