融合注意力机制的CNN-LSTM模型预测蒸发皿蒸发量  

Prediction on pan evaporation by CNN-LSTM model incorporating attention mechanism

作　　者：李少恒 严新军[1,2] 韩克武 王旭 杨怡民 LI Shaoheng;YAN Xinjun;HAN Kewu;WANG Xu;YANG Yimin(College of Water Conservancy and Civil Engineering,Xinjiang Agricultural University,Urumqi 830052,China;Xinjiang Key Laboratory of Hydraulic Engineering Security and Water Disaster Prevention,Urumqi 830052,China)

机构地区：[1]新疆农业大学水利与土木工程学院,新疆乌鲁木齐830052 [2]新疆水利工程安全与水灾害防治重点实验室,新疆乌鲁木齐830052

出　　处：《人民长江》2025年第2期75-81,共7页Yangtze River

基　　金：新疆维吾尔自治区重点研发任务专项项目(2022B03024-3);新疆水利工程安全与水灾害防治重点实验室研究项目(ZDSYS-YJS-2022-09)。

摘　　要：蒸发量数据对于水循环机制理解、水资源规划和农业灌溉管理等领域至关重要。基于新疆吐鲁番地区1973~2022年的逐日气象数据,按各气象因素重要性进行分组,以卷积神经网络(CNN)和长短期记忆神经网络(LSTM)模型为基础模型,融入注意力机制(Attention)以增强模型对关键气象因素的识别和处理能力,构建了一种CNN-LSTM-Attention组合模型来预测蒸发皿蒸发量。将该组合模型与单一模型CNN、LSTM进行对比分析,并采用吐鲁番地区气象站蒸发皿观测数据验证模型的预测精度。研究表明:无论是单一模型还是组合模型,预测精度随输入气象因素增多而提升;组合模型在预测蒸发量方面显著优于传统单一模型,决定系数达到了0.96,相较于传统单一模型CNN、LSTM分别提高了5.4%和6.4%。研究成果可为水资源管理提供数据驱动解决方案。Evaporation data is crucial for understanding the water cycle mechanism,water resource planning,and agricultural irrigation management.Based on the daily meteorological data of Turpan,Xinjiang from 1973 to 2022,meteorological factors were grouped according to their importance.A CNN-LSTM-Attention combined model for the prediction of pan evaporation was constructed by taking the convolutional neural network(CNN)and long short-term memory neural network(LSTM)as the basic models and integrating the attention mechanism to enhance the model′s ability to identify and process key meteorological factors.The combined model was compared with the single models CNN and LSTM,and the prediction accuracy was verified using the observed pan evaporation data from the meteorological stations in Turpan.The research shows that the prediction accuracy of both single models and combined models increases with the augment of input meteorological factors.The combined model significantly outperforms the traditional single models in predicting evaporation,with a coefficient of determination reaching 0.96,which is 5.4%and 6.4%higher than that of the traditional single models CNN and LSTM,respectively.The research results can provide data-driven solutions for water resource management.

关 键 词：蒸发皿蒸发量 卷积神经网络 长短期记忆神经网络 注意力机制 斯皮尔曼相关系数 

分 类 号：TV213.4[水利工程—水文学及水资源] P333.1[天文地球—水文科学]