引入小波分解的Seq2Seq水质多步预测模型研究  

Multi-step prediction model of Seq2Seq water quality with wavelet decomposition

作　　者：白雯睿 杨毅强 李强 BAI Wenrui;YANG Yiqiang;LI Qiang(School of Automation and Information Engineering,Sichuan University of Science and Engineering,Yibin 644000,China)

机构地区：[1]四川轻化工大学自动化与信息工程学院,四川宜宾644000

出　　处：《现代电子技术》2022年第17期100-105,共6页Modern Electronics Technique

基　　金：四川理工学院四川省院士(专家)工作站项目(2018YSGZZ04);自贡市科技局项目(2019YYJC02)。

摘　　要：针对现有水质预测模型对水质多步预测大多采用向量输出的预测模式,忽略了时序预测的输出之间存在的时序联系,导致水质多步预测性能较差的问题,采用小波分解(WD)分解水质数据来提取隐藏的水质特征,然后基于分解所得的序列,建立以长短时记忆(LSTM)网络作为编码器和解码器的序列到序列(Seq2Seq)的预测模型,以期望解决时序预测的输出序列之间存在的依赖性问题。采用珠江流域老口站的溶解氧数据验证模型进行7日预测的效果,实验结果表明,LSTM模型处理该问题的能力要强于传统的MLP及SVR模型,而在LSTM模型的基础上构成的WD-Seq2Seq模型的预测效果进一步提升,溶解氧的7日预测平均MAE仅有0.1471,7日预测平均MSE仅有0.0412,7日预测平均RMSE仅有0.1973,水质类别的7日预测平均准确率达到93.26%。In the existing water quality prediction models,vector output prediction mode is usually adopted for multi-step prediction of water quality,which ignores the time-series relationship among the outputs of time-series prediction,which leads to the poor performance of multi-step prediction of water quality.Wavelet decomposition(WD)is used to decompose the water quality data to extract the hidden water quality features.And then,on the basis of the sequences obtained by decomposition,a sequence-to-sequence(Seq2Seq)prediction model with long and short-term memory(LSTM)network as encoder and decoder is established to eliminate the dependence among the output sequences of time-series prediction.The dissolved oxygen data validation model for Laokou Station in Pearl River Basin was used for seven-day forecast.The experimental results show that the LSTM model has a better ability to eliminate the dependence than the traditional MLP and SVR models.The prediction effect of WD-Seq2Seq model constructed on the basis of the LSTM model is improved further.Its seven-day forecast average MAE of dissolved oxygen is only 0.1471,its seven-day forecast average MSE is only 0.0412,its seven-day forecast average RMSE is only 0.1973,and its seven-day forecast average accuracy of water quality category reaches 93.26%.

关 键 词：小波分解 LSTM模型 Seq2Seq模型 多步预测 时间序列 水质预测 水质指标 溶解氧 

分 类 号：TN911.1-34[电子电信—通信与信息系统]