A multivariate and space-time stochastic weather generator using a latent Gaussian framework

Stochastic weather generators are probabilistic tools used to simulate synthetic time series whose statistics resemble those observed. These tools face challenges in accurately simulating multiple weather variables in space and time, as they require models that can capture complex dependencies between variables and in space-time. We propose a novel spatio-temporal multivariate weather generator, called MSTWeatherGen, that takes advantage of recent developments to model and simulate different weather variables, including temperature, precipitation, wind speed, humidity, and solar radiation, in space and time. Specifically, we use an approach that involves a nonlinear and nonstationary marginal transformation of a multivariate Gaussian random field, characterized by a stationary and nonseparable multivariate spatio-temporal cross-covariance function. To better account for the time-varying nature of weather variables, we divide the time domain into states called weather types. The method is evaluated in the Provence-Alpes-Côte-d'Azur region of France, which is characterized by heterogeneous topography and weather conditions. The evaluation results demonstrate the effectiveness of this new stochastic weather generator in reproducing a wide range of weather statistics, including highly nonlinear indicators such as heat wave or fire weather index.

Obakrim, S., Benoit, L., Allard, D. A multivariate and space-time stochastic weather generator using a latent Gaussian framework. 2024. Link to the Open Access full paper

https://github.com/sobakrim/MSTWeatherGen