Self-clustered GAN for precipitation nowcasting.

Area of Science:

• Meteorology
• Artificial Intelligence
• Machine Learning

Background:

• Precipitation nowcasting models struggle with diverse space-time precipitation patterns due to single latent space limitations.
• Environmental factors introduce non-stationary distributions in precipitation systems, challenging existing models.

Purpose of the Study:

• To develop a novel Generative Adversarial Network (GAN) framework, ClusterCast, for improved precipitation nowcasting.
• To address the limitations of single latent space representations in capturing disparate precipitation distributions.

Main Methods:

• Proposed a GAN framework with a self-clustering approach for precipitation nowcasting.
• Implemented a sub-network for automatic precipitation type labeling and hierarchical feature clustering.
• Utilized self-supervised labels for predicting future radar frames and enabling heterogeneous latent representation.
• Incorporated an ensemble forecast with random perturbations to enhance prediction performance.

Main Results:

• ClusterCast generates non-blurry future radar frames by mitigating mode collapse.
• The method demonstrates robustness across various precipitation scenarios.
• Achieved a 8.9% Mean Squared Error (MSE) on unseen datasets, outperforming four benchmarks for 2-hour predictions.

Conclusions:

• ClusterCast's flexible representation learning accurately models precipitation distribution.
• The self-clustering GAN framework significantly improves precipitation nowcasting accuracy and robustness.
• The proposed method offers a superior alternative to existing benchmarks for short-term precipitation forecasting.