基于CNN-BiLSTM模型的平原型水库洪水预报研究  

Research on Flood Forecasting of Plain Reservoir Based on CNN-BiLSTM Model

作　　者：赵忠峰 王雪妮[1,2] 晋华 郑婕[1] 刘晓东 郭园 ZHAO Zhong-feng;WANG Xue-ni;JIN Hua;ZHENG Jie;LIU Xiao-dong;GUO Yuan(College of Water Resources Science and Engineering,Taiyuan University of Technology,Taiyuan 030024,China;Henan Key Laboratory of Water Resources Conservation and Intensive Utilization in the Yellow River Basin,Zhengzhou 450046,China;PowerChina Northwest Engineering Corporation Limited,Xi’an 710065,China)

机构地区：[1]太原理工大学水利科学与工程学院,山西太原030024 [2]河南省黄河流域水资源节约集约利用重点实验室,河南郑州450046 [3]中国电建集团西北勘测设计院有限公司,陕西西安710065

出　　处：《水电能源科学》2025年第2期10-14,共5页Water Resources and Power

基　　金：山西省科技厅基础研究计划项目(202203021222112);华北水利水电大学河南省黄河流域水资源节约集约利用重点实验室开放研究基金(HAKF202104);国家自然科学基金面上项目(52379018)。

摘　　要：在平原型水库反推入库流量过程中,存在明显的噪声干扰,导致传统的洪水预报方法精度下降。对此,提出一种结合卷积神经网络(CNN)与双向长短期记忆神经网络(BiLSTM)的入库洪水预报模型,该模型采用CNN的卷积层挖掘入库洪水数据中的深层特征信息,并赋予不重要特征较低的权重,以便模型更加专注于对目标任务关键的特征信息。此外,利用BiLSTM处理流量序列中的长期依赖问题,通过其遗忘门有选择性地过滤掉权重较低的特征信息,实现对入库洪水过程的准确预测。最后,基于不同预见期评估所构建模型在安徽省合肥市大房郢水库入库洪水预报中的精准度。结果表明,4 h预见期下CNN-BiLSTM模型在入库洪水预报中具有更高的预报精度,相比BiLSTM模型和新安江(XAJ)模型,其确定性系数(D_(DC))分别提升9.9%、39.0%,均方根误差(R_(RMSE))和相对偏差(B_(BIAS))分别降低34.6%、17.1%和148.6%、20.6%。研究成果可为反推入库流量过程的平原型水库入库洪水预报提供新思路和技术支持。In the process of back-calculating inflow for plain reservoirs,significant noise interference leads to a decrease in the accuracy of traditional flood forecasting methods.To address this,a inflow flood forecasting model combining convolutional neural networks(CNN)and bidirectional long short-term memory(BiLSTM)is proposed.The model utilizes the convolutional layers of CNN to extract deep feature information from the inflow data,assigning lower weights to less important features so that the model can focus more on the key feature information for the target task.Additionally,BiLSTM is employed to address the long-term dependency issue in the flow sequence,selectively filtering out lowweight features through its forget gate,thus enabling accurate prediction of the inflow flood process.Finally,the model's performance in inflow flood forecasting at the Dafangying Reservoir in Hefei,Anhui Province,is evaluated based on different forecasting periods.The results show that the CNN-BiLSTM model achieves higher forecasting accuracy for a 4-hour lead time.Compared to the BiLSTM model and the Xin'Anjiang model,its coefficient of determination(D_(DC))is improved by 9.9%and 39.0%,respectively,while the root mean square error(R_(RMSE))and relative deviation(B_(BIAS))are reduced by 34.6%,17.1%,and 148.6%,20.6%,respectively.The research findings provide new insights and technical support for inflow flood forecasting in plain reservoir.

关 键 词：平原型水库 卷积神经网络 双向长短期记忆神经网络 入库洪水预报 

分 类 号：TV122[水利工程—水文学及水资源] P338[天文地球—水文科学]