Application and accuracy evaluation of a multi-temporal integrated AFNO-based rainfall prediction model for river basins in Kyushu

Abstract

In recent years, heavy rainfall from line-shaped rainbands has increased, highlighting the need for accurate precipitation forecasts up to 6 hours ahead. We develop a multi-temporal AFNO-based model integrating precipitation and precipitable water vapor (PWV) into channel-wise inputs.Evaluation over the Chikugo and Kuma River basins shows improved accuracy up to 6 hours ahead, particularly when boundary effects are considered and PWV predictions are included. Comparisons with NowCastNet and the JMA short-term forecast demonstrate comparable performance for intense rainfall prediction.