Abstract
Nowcasting is an observation-based method that uses the current state of the atmosphere to forecast future weather conditions over several hours. Recent studies have shown the promising potential of using deep learning models for precipitation nowcasting. In this paper, novel deep generative models are proposed for precipitation nowcasting. These models are equipped with extreme-value losses to more reliably predict extreme precipitation events. The proposed deep generative model contains a Vector Quantization Generative Adversarial Network and a Transformer ("VQGAN + Transformer"). For enhanced modeling and forecasting of extreme events, Extreme Value Loss (EVL) is incorporated in the autoregressive Transformer. The numerical results show that the proposed model achieves comparable performance with the state-of-the-art conventional nowcasting method PySTEPS for predicting nominal values. By incorporating an EVL, the proposed model yields more accurate nowcasting of extreme precipitation.
Introduction: The Challenge of Extreme Nowcasting
Predicting weather conditions within a 0 to 6-hour window, known as "nowcasting," is a critical component of early flood warning systems. Traditional methods typically rely on "optical flow" techniques to track the movement of precipitation cells. While deep learning has recently emerged as a powerful alternative, standard models face two significant hurdles: the "blurriness trap" and the "imbalance problem."
Models trained on standard loss functions, such as Mean Squared Error (MSE), often produce "average" looking images. This results in blurry forecasts that fail to capture the sharp intensity and structure of real storms. Furthermore, extreme rainfall is statistically rare; in many datasets, over 90% of observations record less than 1mm of rain. Standard models become highly proficient at predicting "no rain" but struggle to identify the rare, high-impact events that lead to flooding.
Methodology: The VQGAN + Transformer Framework
To address these challenges, we propose a two-stage generative architecture: the VQGAN + Transformer. This approach moves away from predicting raw pixel values, which often leads to smoothing, and instead focuses on learning a discrete representation of weather patterns.
Stage 1: VQGAN
The Vector Quantized Generative Adversarial Network (VQGAN) learns a "codebook" of common precipitation patterns. It compresses complex radar maps into a sequence of discrete tokens, effectively creating a visual vocabulary for weather. This discretization helps the model maintain sharp boundaries and realistic storm structures.
Stage 2: 3DNA Transformer
An autoregressive Transformer, utilizing a 3D Neural Attention (3DNA) layer, predicts the next sequence of tokens. This allows the model to capture both complex spatial patterns and their temporal evolution over time.
Extreme Value Loss (EVL)
The core innovation is the integration of Extreme Value Theory (EVT) into the loss function. Rather than relying solely on standard Cross-Entropy, we incorporate an Extreme Value Loss (EVL) based on the Generalized Pareto Distribution. This mathematically incentivizes the model to prioritize the "tail" of the rainfall distribution, ensuring that rare, high-intensity events are not ignored as statistical noise.
Data and Study Area
The study utilized a high-resolution dataset from the Royal Netherlands Meteorological Institute (KNMI), covering the period from 2008 to 2021. The data features a 1 km spatial resolution and 5-minute temporal resolution. Validation was performed across 12 specific Dutch lowland river catchments, providing a hydrologically relevant assessment of the model's performance.
Key Results and Discussion
The proposed model was compared against PySTEPS, the current state-of-the-art conventional nowcasting framework. For standard, nominal rainfall, the deep learning model achieved performance comparable to PySTEPS (PCC of ~0.21). However, the model's true strength appeared when evaluating extreme events.
As shown in the ROC curves above, the CE + EVL (Cross Entropy + Extreme Value Loss) model significantly outperforms the baseline methods for the top 1% of events. At a fixed Hit Rate (HR) of 0.8, the proposed model maintained a lower False Alarm Rate (0.31 vs. 0.36 for PySTEPS) and a higher Critical Success Index (0.378 vs. 0.353). Visually, the VQGAN approach successfully avoided the blurry "blobs" typical of older deep learning models, maintaining sharper and more realistic storm structures.
Conclusion
By combining generative adversarial networks with autoregressive transformers and a specialized extreme value loss, we have developed a nowcasting model that excels where traditional deep learning often fails: the prediction of rare, high-intensity precipitation. While the model is currently purely data-driven, future work will explore the integration of physical and thermodynamic constraints to further refine the accuracy and extend the reliable lead time of these forecasts.
Summary updated