Implementing generative adversarial network (GAN) as a data-driven multi-site stochastic weather generator for flood frequency estimation

Precipitation is a key driving factor of hydrologic modeling for impact studies. However, there are challenges due to limited long-term data availability and complex parameterizations of existing stochastic weather generators (SWGs) due to spatiotemporal uncertainty. We introduced state-of-the-art Generative Adversarial Network (GAN) as a data-driven multi-site SWG and synthesized extensive hourly precipitation over 30 years at 14 stations. These samples were then fed into an hourly-calibrated SWAT model for streamflow generation. Results showed that the well-trained GAN improved rainfall data by accurately representing spatiotemporal distribution of raw data rather than simply replicating its statistical characteristics. GAN also helped display authentic spatial correlation patterns of extreme rainfall events well. We concluded that GAN offers a superior spatiotemporal distribution of raw data compared to conventional methods, thus enhancing the reliability of flood frequency evaluations. •Generative Adversarial Networks promise to replace traditional weather generators.•Deep learning was used to bypass the arduous work of defining complex parameters.•Deep learning can detect intricate patterns in a multidimensional distribution.•Generative Adversarial Network captures spatial correlation of extreme events.