水力发电2025，Vol.51Issue(3)：12-21,56,11.
考虑水文机理的深度学习径流预测模型及其可解释性
Deep Learning-based Runoff Prediction Models Incorporating Hydrological Mechanisms and an Analysis of Model Interpretability
摘要
Abstract
To enhance the accuracy of daily runoff prediction and bolster the interpretability of deep learning models,the physical processes of hydrological models can be incorporated as inputs to these models.The output data from various modules of Xin'anjiang model are utilized as additional feature inputs for two deep learning models of GRU and GRU-Seq2seq-Attention,respectively,to study the influences of different input features and model structures on runoff prediction results,and the interpretability of deep learning runoff prediction models is analyzed by the integral gradient method.The results show that,(a)introducing different input features into the deep learning model can significantly improve the accuracy of runoff prediction,in which,the GRU-Seq2seq Attention model considering simulated runoff has the best prediction accuracy;and(b)the analysis of global and local interpretability reveals the preferences of different models in capturing key features in different traffic scenarios,which provides important guidance for the subsequent optimization and improvement of the models.关键词
径流预测/Seq2seq模型/深度学习模型/耦合模型/可解释性Key words
runoff prediction/Seq2seq model/deep learning model/coupling model/interpretability分类
地球科学引用本文复制引用
姚泰伦,肖培伟,陆宝宏,熊子云,王淼,王冰冰..考虑水文机理的深度学习径流预测模型及其可解释性[J].水力发电,2025,51(3):12-21,56,11.基金项目
国家自然科学基金联合基金项目(U2240218) （U2240218）
金沙江旭龙水电站建设期影响区域环境水体监测及评估项目(XL-FW-2023-012) （XL-FW-2023-012）
