Modeling precipitation evolution with global deterministic motion and local stochastic variations improves prediction accuracy.
GPTCast is a generative deep learning method that leverages a GPT model and a specialized spatial tokenizer to produce realistic and accurate ensemble forecasts of radar-based precipitation.
A novel deep learning model leveraging multi-source meteorological data and temporal attention mechanisms surpasses existing operational methods in predicting rainfall in Denmark up to 8 hours in advance.
This paper introduces DTCA, a novel Transformer-based diffusion model for precipitation nowcasting, which leverages causal attention and channel-to-batch shifting to enhance the capture of spatiotemporal dependencies and improve prediction accuracy.
The Multi-Task Latent Diffusion Model (MTLDM) improves the accuracy of short-term precipitation forecasting, particularly for extreme events, by decomposing radar images into sub-images based on precipitation intensity and predicting each component separately.