Precipitation Nowcasting Using Physics Informed Discriminator Generative Models (2406.10108v1)

Abstract: Nowcasting leverages real-time atmospheric conditions to forecast weather over short periods. State-of-the-art models, including PySTEPS, encounter difficulties in accurately forecasting extreme weather events because of their unpredictable distribution patterns. In this study, we design a physics-informed neural network to perform precipitation nowcasting using the precipitation and meteorological data from the Royal Netherlands Meteorological Institute (KNMI). This model draws inspiration from the novel Physics-Informed Discriminator GAN (PID-GAN) formulation, directly integrating physics-based supervision within the adversarial learning framework. The proposed model adopts a GAN structure, featuring a Vector Quantization Generative Adversarial Network (VQ-GAN) and a Transformer as the generator, with a temporal discriminator serving as the discriminator. Our findings demonstrate that the PID-GAN model outperforms numerical and SOTA deep generative models in terms of precipitation nowcasting downstream metrics.

• The paper introduces PID-GAN, a physics-informed deep generative model combining VQ-GAN, Transformer, and a temporal discriminator for precipitation nowcasting.
• By incorporating the moisture conservation equation, PID-GAN enhances the physical consistency of precipitation forecasts compared to models lacking such constraints.
• Empirical results show PID-GAN outperforms state-of-the-art methods across multiple metrics, particularly in detecting extreme precipitation events, highlighting the benefit of physics integration.

This paper introduces a physics-informed deep generative model for precipitation nowcasting, leveraging a PID-GAN to integrate physical priors of precipitation for accurate and physically consistent forecasts.

• The authors propose a PID-GAN architecture comprising a VQ-GAN and an autoregressive Transformer as the generator, coupled with a temporal discriminator, to effectively model spatio-temporal precipitation dynamics.
• By incorporating the moisture conservation equation $$\mathcal{R}_q = -\frac{\partial q}{\partial t} -u_{10} \frac{\partial q}{\partial x} - v_{10} \frac{\partial q}{\partial y} -u_{100} \frac{\partial q}{\partial x} - v_{100} \frac{\partial q}{\partial y} + ET - P$$, where $q$ is specific humidity, $u$ and $v$ are wind components at different altitudes, $ET$ is evapotranspiration, and $P$ is precipitation, the model enhances physical consistency in precipitation forecasts.
• Empirical results demonstrate that the PID-GAN model outperforms state-of-the-art baselines, including PySTEPS and NowcastingGPT, in terms of MSE, PCC, CSI, FAR, FSS, and AUC, particularly in detecting extreme precipitation events, evidenced by a 6.17% AUC decrease when physical constraints are removed.