基于注意力机制与LSTM-CCN的月降水量预测
Monthly precipitation prediction based on attention mechanism and LSTM-CCN
周祥 1张世明 2苏林鹏 3张守平1
作者信息
- 1. 重庆水利电力职业技术学院,重庆 402160;水库安全及水环境大数据重庆市高校工程中心,重庆 402160
- 2. 长江水利委员会水文局长江上游水文水资源勘测局,重庆 400020
- 3. 重庆市渝西水利电力勘测设计院有限公司,重庆 402160
- 折叠
摘要
针对现有月降水量预测方法预测准确性不高的问题,提出一种基于注意力机制与LSTM-CCN的月降水量预测方法.首先,利用长短时记忆神经网络(long short-term memory neural network,LSTM)提取气象数据在时间维度的特征分布,从时间相关性方面捕获相邻时间段或长距离气象数据段中的统计分布;其次,利用因果卷积神经网络(causal convolutional network,CCN)将气象数据映射到空间维度,深层次地从空间维度捕获气象数据在空间中的特征统计分布;再次,以并联的方式将时间和空间特征作为交叉注意力网络的输入,构造融合的时空特征;最后,以长短时记忆神经网络构造解码器,并将融合的时空特征作为解码器的输入,预测的月降水量作为输出.选取河南省新乡市2001～2017 年数据集进行测试,结果表明:所提出方法的均方根误差仅为13.08 mm,相比主流方法具有更低的预测误差.研究成果可为提高气象预测的准确性和实用性提供参考.
Abstract
To address the issue of low accuracy in existing monthly precipitation prediction methods,an attention mechanism and LSTM-CCN for the monthly precipitation prediction method were proposed.Firstly,the long short-term memory neural net-work(LSTM)was used to extract the temporal feature distribution of meteorological data,capturing the statistical distribution in adjacent or long-distance meteorological data segments from a temporal correlation perspective.Secondly,the causal convolution-al network(CCN)projected meteorological data to the spatial dimension,deeply capturing the statistical distribution of spatial features of meteorological data.Thirdly,the time and space features were input into the cross-attention network in parallel,con-structing a fused spatiotemporal feature.Finally,a decoder constructed with the long short-term memory neural network took the fused spatiotemporal feature as input,and the predicted monthly precipitation served as the output.The test was carried out on the data set from Xinxiang City,Henan Province from 2001 to 2017.The results showed that the proposed method's root mean square error was only 13.08 mm,demonstrating lower prediction errors compared to mainstream methods.The introduction of this work contributes to enhancing the accuracy and practicality of meteorological predictions.
关键词
月降水量预测/多层注意力机制/因果卷积神经网络/长短时记忆神经网络Key words
monthly precipitation prediction/multi-layer attention mechanism/causal convolutional neural network/long short-term memory neural network引用本文复制引用
基金项目
重庆市技术创新与应用发展专项重点项目(CSTB2022TIAD-KPX0132)
出版年
2024
国家科技期刊平台
NSTL
万方数据