Performance enhancement of deep learning model with attention mechanism and FCN model in flood forecasting

Abstract

Accurate and timely inflow flood forecasting is a critical foundation for the flood operation of multi-reservoir systems. To address the issue of low accuracy in long lead-time and extreme event flood forecasting with existing deep learning models, this study employed a Gated Recurrent Unit (GRU) as the base model and integrated the Attention mechanism and one-dimensional fully convolutional network (FCN) module to enhance its performance. The upper reaches of the Luohe River basin in the middle and lower Yellow River basin in China were chosen as the study area, utilizing observed data from 14 rain gauge stations, 1 evaporation station, and 1 hydrological station from 2013 to 2021 to build the dataset. GRU, GRU-FCN, Attention-GRU, and Attention-GRU-FCN were applied to flood events and daily streamflow forecasting. In addition, Informer was introduced and compared with other models. The results showed that the Attention mechanism enhanced GRU’s ability to predict extreme flood events while achieving more stable forecasting results, thereby improving the model’s performance for long lead times. The FCN module further strengthened the performance of GRU and Attention-GRU. Among the four GRU-based models, Attention-GRU-FCN demonstrated the best performance in extreme flood forecasting and long lead-time predictions, with the smallest peak timing error. Informer exhibited a significant advantage in long lead-time predictions but had lower accuracy in peak flood forecasting compared to Attention-GRU-FCN.