Ensemble precipitation nowcasting by combination of generative and transformer deep learning models

This work introduces a novel deep-learning method for generating realistic ensembles nowcast of radar-based precipitation at a five-minute time resolution for the next 60 minutes and longer.

The proposed method is composed of a combination of two models: the first model is trained to compress and decompress the spatial domain into and from a discrete representation (tokens), while the second model evolves the compressed representation over time. Specifically, the compression and decompression model is based on a combination of a Quantized Variational Autoencoder with a Generative Adversarial Network, while the prediction over time leverages a Generative Pretrained Transformer (GPT) architecture.

This separation of concerns (discretized spatial compression/decompression and temporal extrapolation) adds several desirable features not present in more commonly used deep learning methods based on recurrent/convolutional deep learning architectures: 

• transformer output probabilities can be leveraged to generate ensemble/probabilistic forecasts (without the need of injecting noise)
• the discretized spatial representation can be used to characterize each token, adding interpretability and explainability to the model
• the combination of transformer probabilities and token characterization can be used at inference time for forecasts conditioning based on external factors (e.g. NWP forecast output)

The presented architecture is trained and tested on a 7-year radar dataset of reflectivity composites of the Emilia-Romagna Region, Italy. The method is then applied at two different scales: regional, over Emilia-Romagna, and national, on the entire Italian domain, showing the adaptability of the approach to multiple spatial domains. We will present the performance of this model for both deterministic and ensemble settings by comparing it with respect to other commonly used extrapolation and deep learning methods.