1:45 PM - 2:00 PM
[ATT35-01] Real-time nowcasting of "Guerrilla" rainstorms: Demonstration at Osaka Expo 2025
★Invited Papers
Keywords:Downpour, nowcasting, neural network, Guerrilla rainfall
Guerrilla rainstorms are sudden, localized downpours that have been occurring with increasing frequency in Japan. These intense storms can cause significant infrastructure damage, making them a serious societal issue. Due to their sudden onset, localized nature, and vertical development at high altitudes, they remain difficult to predict, even with lead times as short as 10 minutes.
To address this challenge, Multi-Parameter Phased Array Weather Radars (MP-PAWR) have been developed to provide high-resolution atmospheric observations. Initially tested over Saitama in 2018, two new versions of the instrument have been deployed in Osaka and Kobe since the summer of 2024.
Leveraging the dense four-dimensional (4D) observations in space and time, novel radar extrapolation methods have been developed for guerrilla rainstorm nowcasting, i.e., high-resolution, very-short-term forecasting. At NICT, a 4D Artificial Neural Network (ANN) has been developed and trained on Saitama’s 2020 data for 10-minute lead-time nowcasting. This model, which combines Long Short-Term Memory (LSTM) units with 3D spatial convolutions and an adversarial training technique, extrapolates 3D radar maps into future time steps. It has been integrated into a lightweight real-time system that will be demonstrated at the Expo 2025 Osaka "Future Society Showcase Project: Future Life Experience." Although the model was trained using 2020 Saitama data, initial results show promising performance when applied to the new MP-PAWR instruments of Kobe and Suita.
In this presentation, we will introduce the system and compare its nowcasting performance with a TREC-based real-time system developed by RIKEN, which will also be used at the Osaka Exposition.
To address this challenge, Multi-Parameter Phased Array Weather Radars (MP-PAWR) have been developed to provide high-resolution atmospheric observations. Initially tested over Saitama in 2018, two new versions of the instrument have been deployed in Osaka and Kobe since the summer of 2024.
Leveraging the dense four-dimensional (4D) observations in space and time, novel radar extrapolation methods have been developed for guerrilla rainstorm nowcasting, i.e., high-resolution, very-short-term forecasting. At NICT, a 4D Artificial Neural Network (ANN) has been developed and trained on Saitama’s 2020 data for 10-minute lead-time nowcasting. This model, which combines Long Short-Term Memory (LSTM) units with 3D spatial convolutions and an adversarial training technique, extrapolates 3D radar maps into future time steps. It has been integrated into a lightweight real-time system that will be demonstrated at the Expo 2025 Osaka "Future Society Showcase Project: Future Life Experience." Although the model was trained using 2020 Saitama data, initial results show promising performance when applied to the new MP-PAWR instruments of Kobe and Suita.
In this presentation, we will introduce the system and compare its nowcasting performance with a TREC-based real-time system developed by RIKEN, which will also be used at the Osaka Exposition.