Attention-enhanced sequential deep learning models for short-term typhoon rainfall prediction over eastern Taiwan

Abstract

Taiwan is located in the western North Pacific, where typhoon-induced heavy rainfall frequently challenges disaster prevention and decision-making. To improve short-term (1-6 h) rainfall forecasting during typhoon landfall, this study developed a deep learning framework for eastern Taiwan using historical typhoon records and ground-based meteorological observations. Six models were compared: Transformer Encoder model (TransEnc), LSTM, GRU, attention-enhanced LSTM (MSA-LSTM), attention-enhanced GRU (MSA-GRU), and an application-oriented hybrid Encode-Decode-Attention-Recurrent (EDAR) framework. At t + 1, all models broadly captured the main rainfall evolution. As forecast horizon increased, RMSE and MAE generally rose, whereas NSE and correlation declined, indicating recursive error accumulation. MSA-GRU showed the strongest overall multi-step performance, with the highest average NSE and correlation and significantly lower RMSE than TransEnc, LSTM, GRU, and MSA-LSTM (p < 0.05). EDAR achieved the lowest mean MAE and remained competitive in the supplementary Keelung analysis. Hualien served as the primary development site, and a supplementary analysis was also conducted at Keelung. Overall, attention-enhanced recurrent models provided more stable multi-step forecasts, with potential for real-time typhoon heavy-rainfall early warning.