Development of AI-Based Nowcasting Model for Precipitation Using SimVP: Simpler yet Better Video Prediction Method and RADAR Data

In the field of weather forecasting, nowcasting precipitation within 6-hours timeframe plays a crucial role in protecting public safety and property from various types of rapidly developing, hard-to-predict precipitation events. Recently, significant global progress has been made in developing AI-based weather prediction techniques, which, in some aspects, surpass traditional numerical weather prediction (NWP) models, with their performance reaching levels comparable to those of NWP models and radar observations. AI-based weather prediction leverages decades of accumulated meteorological data for training and testing, enabling the construction of sophisticated models. This study focuses on developing a nowcasting model for precipitation using SimVP, a video prediction method optimized for meteorological data, combined with Generative Adversarial Networks (GAN). The trained SimVP generates temporal prediction images, which are then evaluated by the GAN's discriminator by comparing them with actual observation images, thereby predicting the movement of precipitation cells with high accuracy. The SimVP model demonstrates superior performance in video prediction tasks, outperforming vision transformer models by utilizing convolution operations alone. For model development, radar composite reflectivity and accumulated precipitation data for the years 2018, 2019, 2021, and 2022 were used for training, while data from 2020 were used for testing. The developed model performs 6-hour predictions at 10-minute intervals, estimating the probability of precipitation in different intensity ranges and ultimately predicting precipitation amounts. The newly improved model has shown significant performance enhancements, with FSS, F1, and CSI metrics improving by 40.7%, 28.7%, and 32.4%, respectively, compared to the 2022 model. Additionally, the model was evaluated using real-time radar data and analyzed with Structure-Amplitude-Location (SAL) and Power Spectral Density (PSD) methods, confirming high spatiotemporal accuracy. Fine-tuning the model with the AWS rain gauge data has also enabled predictions for AWS (Automatic Weather System) sites. As a result, the model can provide predicted radar images and precipitation data for the Korean Peninsula at 10-minute intervals over a 6-hour period, supporting forecasters in nowcasting. This advancement offers faster and more accurate predictions than traditional nowcasting methods, making it an effective tool for responding to the increasingly frequent and sudden extreme weather driven by climate change.

Key words: RADAR, AI, SimVP, Generative Adversarial Networks, Precipitation

This work was funded by the Korea Meteorological Administration Research and

Development Program "Developing AI Technology for Weather Forecasting" under Grant (KMA2021-00121).