摘要
为有效提取河道径流时间序列信息特征,提高河道汇流过程模拟预测的非线性拟合能力,构建一种融合双向长短期记忆网络(Bi-LSTM)、多头注意力机制(Multi-Head Attention)、前馈神经网络(FFNN)的河道汇流预测模型(MABLFN)。为验证MABLFN模型有效性,以永定河山峡段典型站点实测数据开展实例验证,并将预测结果与单一的LSTM、Bi-LSTM模型和具有物理机制的MIKE11模型预测结果进行对比分析,评估模型不同预报时长径流过程预测性能。结果表明:MABLFN模型能够较好地预测河道径流,MABLFN模型相比于LSTM模型、Bi-LSTM模型和MIKE11模型的RMSE降低了1%~52%,NSE提高了8%~9%;在计算效率方面MABLFN模型相比于LSTM模型、Bi-LSTM模型计算耗时由0.26 s增加至1.2 s,相比于MIKE11模型(360 s)计算耗时明显降低。
To effectively extract the characteristics from river runoff time series and improve the nonlinear fitting capability of river routing simulations,this study constructed a river routing model(MABLFN)that integrated Bi-directional Long Short-Term Memory networks(Bi-LSTM),Multi-Head Attention mechanisms,and Feed Forward Neural Networks(FFNN).Empirical validation was conducted using measured data from typical stations in the mountainous section of the Yongding River to validate the effectiveness of the MABLFN model.The prediction results were compared with those from single LSTM,Bi-LSTM models,and the MIKE11 model with physical mechanisms,to evaluate the predictive performance of the model for runoff processes over different forecast periods.The results indicated that the MABLFN model is able to predict runoff processes.Compared to LSTM,single Bi-LSTM,and MIKE11,the RMSE of MABLFN is reduced by 1%to 52%,the NSE increased by 8%to 9%.Considering computation time,the MABLFN model increased the computation time from 0.26 s to 1.2 s compared to LSTM and Bi-LSTM,while reducing the computation time compared to MIKE11(360 s).
作者
程帅
张娟
李晓琳
杨默远
沈建明
CHENG Shuai;ZHANG Juan;LI Xiaolin;YANG Moyuan;SHEN Jianming(Beijing Water Science and Technology Institute,Beijing 100048,China;Capital Normal University,Beijing 100089,China)
出处
《水文》
北大核心
2025年第2期80-87,共8页
Journal of China Hydrology
基金
北京市自然科学基金资助项目(8232032)
国家自然科学基金资助项目(52209005)。
关键词
河道汇流演算
双向长短期记忆网络
多头注意力机制
深度学习
river routing
Bidirectional Long Short Term Memory Network
Multi-Head Attention mechanism
deep learning
