Thursday, 31 August 2023
Boundary Waters (Hyatt Regency Minneapolis)
The latest generative models can be grouped into likelihood-based models, score-based models, and implicit models according to the method of representing the probability distribution. In particular, implicit models give up tractability and focus on flexibility to generate high-quality results, among the trade-off of generative models. One of the outstanding implicit models is generative adversarial networks (GANs), in which various mutations exist. GANs can estimate complicated distributions, but the instability of the adversarial training that makes it possible is notorious. The difficulty of adversarial training increases even more when dealing with real-world data with unrefined distributions. For this reason, prominent previous models using GANs for radar-based precipitation nowcasting did not consist of regular GANs alone. The complex network architectures with convolutional long short-term memory (ConvLSTM) or convolutional gated recurrent unit (ConvGRU) were forced. To be free from these constraints, we introduce paired complementary temporal cycle-consistent adversarial networks (PCT-CycleGAN) for radar-based precipitation nowcasting, inspired by cycle-consistent adversarial networks (CycleGAN). PCT-CycleGAN does not use complex network architectures and guarantees temporal causality by itself without ConvLSTM or ConvGRU. Nevertheless, PCT-CycleGAN outperforms quantitative precipitation forecast (QPF) models and previous artificial intelligence (AI) nowcasting models during a lead time of two hours. Our key contributions to PCT-CycleGAN are demonstrated with ablation studies.

