Towards an Operational Precipitation Nowcasting Model with Memory-Efficient Bidirectional Transformers

Precipitation nowcasting is crucial for preventing casualties and economic losses caused by heavy rainfall. While deep learning models like the U-Net convolutional neural network and generative adversarial network have been proposed for this task, their performance remains limited. A promising state-of-the-art technique, the autoregressive transformer, has demonstrated excellent generative modeling capabilities. However, its slow inference time makes it unsuitable for precipitation nowcasting. In this talk, we introduce the results of precipitation nowcasting with 0-6-h lead times using a generative model called Memory-Efficient Bidirectional Transformers (MeBT). Unlike the autoregressive transformer, MeBT offers fast inference time, making it suitable for operational precipitation nowcasting. We conducted comprehensive evaluations of the MeBT model's performance, comparing it to strong traditional baselines such as the U-Net convolutional neural network (CNN), the generative adversarial network, the optical flow model, and the Eulerian Persistence. The MeBT model outperforms these baselines in predicting precipitation events occurring throughout all four seasons, as it provides more accurate predictions for newly developed and/or decayed precipitation fields. Furthermore, we assessed the operational feasibility of the MeBT model through the following evaluations: (1) forecast skill evaluation by comparing it with the current operational models employed by the Korean Meteorological Administration (KMA), such as KLAPS (Korea Local Analysis and Prediction System) and MAPLE (McGill Algorithm for Precipitation nowcasting by Lagrangian Extrapolation), and (2) systematic evaluation by professional forecasters in KMA. According to the evaluations, the MeBT model generally provides superior predictions, and the predicted fields exhibit temporal consistency at 0-6-hour lead times. We believe the MeBT model could enhance the accuracy of the operational nowcasting model.