A Novel Framework for Nonparametric Rainfall Generator Based on Deep Convolutional Wasserstein Generative Networks (DC-WGANs)

Abstract A novel rainfall generator based on Deep convolutional Wasserstein Generative Adversarial Networks (DC-WGANs) is implemented to generate spatial-temporal hourly rainfall for the Kelantan River basin in Malaysia. DC-WGAN as a deep generative model consists of two deep neural networks, generator and discriminator, which compete against each other in a min-max game. The final output of this contest is a generator capable of generating unlimited realistic samples from the dataset. The novelty of our generator is converting time series rainfall data of multi-station to images, then building realistic images based on them, and finally converting back to multivariate time series rainfall. Models' outcomes were reasonable in terms of occurrence, extremes, and amount of rainfall, in hourly, daily, and yearly time resolutions. The model's ability to cover a wide range of rainfall variability can be used in uncertainty analysis and dataset augmentation. In our limited number of stations, a high correlation between synthetic data and original data was observed, ranging R from 0.97 to 0.99; also, despite other nonparametric methods, the proposed generator was able to generate extreme events in good quality and quantity. The conclusion is drawn from a case study in Malaysia but can be generalized worldwide by using global data.