基于改进iTransformer的多维特征河流水质预测方法研究  

Method of multi⁃dimensional feature river water quality prediction based on improved iTransformer

作　　者：樊力震 董建刚 李俊俊 FAN Lizhen;DONG Jiangang;LI Junjun(School of Software Engineering,Xinjiang University,Urumqi 830091,China)

机构地区：[1]新疆大学软件学院,新疆乌鲁木齐830091

出　　处：《现代电子技术》2025年第8期179-186,共8页Modern Electronics Technique

摘　　要：水质预测是水资源生态管理的重要组成部分。水质数据易受环境影响,随着时间、随机事件、自然条件变化等因素呈现出非平稳性和非线性的特性,使得水质时序依赖较为复杂,其规律难以捕捉。为更准确地提取水质时序规律,并使其具备一定的泛化性,提出一种基于改进iTransformer的多维特征水质预测模型——GF-iTransformer。针对水质数据中的复杂噪声问题,引入一维高斯-拉普拉斯滤波器对水质时序数据进行降噪。为更好地挖掘水质数据中隐含的频域信息,加入频率增强通道注意力机制,利用基于离散余弦变换(DCT)的频率信息提取方法,从本质上避免了基于傅里叶变换(FT)造成的吉布斯现象,并相对减少了计算量,得到了更好的预测性能。在3个不同的公共数据集(ETTh1、ETTh2、ETTm2)和两个河流数据集(yihe、luohe)上进行验证,结果表明,相较于TimesNet、ETSformer、DLinear等6个现有主流时序预测模型,文中所提GF-iTransformer模型都展现出了较好的预测精度,证明了该模型的有效性。Water quality prediction is an important part of water resources ecological management.Water quality data is easily affected by the environment,and presents non-stationarity and nonlinear characteristics with the change of time,random events,natural conditions and other factors,which makes the time series dependence of water quality more complicated and its rules difficult to capture.In order to extract the water quality time series law more accurately and make it have a certain generalization,a multi-dimensional feature water quality prediction model based on improved iTransformer,GF-iTransformer,is proposed.In allusion to the complex noise problem in water quality data,a one-dimensional Gauss-Laplace filter is introduced to reduce the noise of water quality time series data.In order to better mine the hidden frequency domain information in water quality data,a frequency-enhanced channel attention mechanism is added,and the frequency information extraction method based on discrete cosine transform(DCT)is used to essentially avoid the Gibbs phenomenon caused by Fourier transform(FT),and reduce the calculation amount,resulting in better prediction performance.The experimental verification is performed on three different public data sets(ETTh1,ETTh2,ETTm2)and two river data sets(yihe,luohe).The results show that in comparison with six existing mainstream time series prediction models such as TimesNet,ETSformer,and DLinear,the proposed GF-iTransformer can model show better prediction accuracy,which proves the effectiveness of GF-iTransformer.

关 键 词：水质预测 多维特征 iTransformer模型 高斯-拉普拉斯滤波器 注意力机制 离散余弦变换 

分 类 号：TN713-34[电子电信—电路与系统] TV213.4[水利工程—水文学及水资源] X522[环境科学与工程—环境工程]