Assessing the Impact of Rainfall Nowcasts on an Encoder-Decoder LSTM Model for Short-Term Flash Flood Prediction

作     者：Mhedhbi, Rim Erechtchoukova, Marina G. 

作者机构：York Univ Sch Informat Technol Fac Liberal Arts & Profess Studies 4700 Keele St Toronto ON M3J 1P3 Canada 

出 版 物：《WATER RESOURCES MANAGEMENT》 (Water Resour. Manage.)

年 卷 期：2025年第39卷第4期

页      面：1623-1638页

核心收录：

学科分类：08[工学] 0815[工学-水利工程] 0814[工学-土木工程] 

基　　金：The authors declare that no funds  grants  or other support were received during the preparation of this manuscript 

主　　题：LSTM Model Encoder-Decoder Architecture Flash Flood Prediction Rainfall Nowcasting Forecast Uncertainty 

摘      要：Flash floods pose significant threats as immediate and highly destructive natural hazards. Extending the forecast horizon of flash flood prediction models has been a key objective to enable timely warning or other mitigating measures. The integration of precipitation predictions into data-driven flash flood models remains unexplored. In this study, we propose an Encoder-Decoder LSTM-based model architecture for short-term flash flood prediction, which incorporates short-term rainfall forecasts and evaluates the influence of the associated uncertainty on these predictions. We conducted three sets of experiments to predict flash flood occurrences within a watershed with a 30-minute response time. The first set employed a baseline LSTM model without rainfall forecast integration. The second one utilized a proposed encoder-decoder LSTM model that incorporated accurate rainfall forecasts. Lastly, the third set of experiments introduced errors into the rainfall forecasts to evaluate the impact of forecast uncertainty on flood prediction. Computational experiments demonstrate that incorporating accurate rainfall nowcasts significantly enhances flash flood predictability, with F1-score improvements ranging from 10 to 60%, depending on the hydrological year. Furthermore, even when errors in rainfall magnitude and timing were introduced, overall the proposed framework outperformed models that did not use rainfall forecasts, delivering reliable predictions for up to two hours.