关键词: 洪水预报；注意力机制；神经网络；深度学习模型；可解释性；向家坝水库；三峡水库；长江流域

Abstract: For improving accuracy of flood forecasting in the Xiangjiaba - Three Gorges Reservoir interval basin and exploring the interpretability of artificial intelligence (AI) deep learning approach, we coupled the feature - temporal dual attention mechanism (DA) and the recursive encoding - decoding process (RED) into the long short - term memory (LSTM) neural network, and constructed a DA - LSTM - RED model. Flood forecasting with 1~7d forecast periods in the Xiangjiaba to Three Gorges Reservoir interval basin was conducted and compared with the LSTM - RED model. The results show that two AI deep learning models have high forecasting accuracy in training and verification periods; the performance of DA - LSTM - RED model is better than LSTM - RED model as the forecast periods prolong, and the Nash efficient coefficient and relative error in 7d forecast periods during validation period are 0.94 and -0.48%, respectively. The proposed DA - LSTM - RED model can identify input variables with high correlation to the target output, which not only enhances the model forecasting ability but also improves the interpretability of the machine deep learning model to a certain extent. This offers a new technical approach for flood simulation and forecasting.

Key words: flood forecasting; attention mechanism; neural network; deep learning model; interpretability; Xiangjiaba Reservoir; Three Gorges Reservoir; Changjiang River Basin