摘要
本文以地表水氨氮因子作为水质分析的研究对象,将系统降噪方法(Seasonal-HybridExtreme Studentized Deviate test,S-H-ESD)、经验模态分解(Empirical Mode Decomposition,EMD)方法和多模态输入长短期记忆模型(Multi-Modal Long Short-Term Memory,Multi-Modal-LSTM)相结合,构建了一个多尺度模态组合预测模型,EMD-Multi-ModalLSTM。在模型构建过程中,首先通过S-H-ESD算法对原始波动数据进行系统性降噪;其次,对降噪后的序列采用EMD分解为不同特征尺度的本征模态分量(IMF)和一个趋势项。在此基础上,对各分量再分别结合其他相关的序列信息,单独构建Multi-Modal-LSTM模型,并进一步通过集成各预测分量获得整体氨氮序列预测值。以珠三角地区两种不同水质断面氨氮浓度为例进行实证分析,结果表明本文方法相比于传统机器学习算法及深度学习算法具有更高的模型预测精度,且对数据波动较大的氨氮序列及高浓度时刻预测效果提升明显,预测性能更加稳定。
In this paper,the ammonia nitrogen factor of surface water was taken as the research object.Combing the system denoising method(S-H-ESD),empirical mode decomposition(EMD) with MultiModel-LSTM to construct a multi-scale-modal combined prediction model,EMD-Multi-Modal-LSTM.During the model building process,the S-H-ESD algorithm is used to systematically reduce the noise of the original fluctuation data.Then the denoised sequence is decomposed by EMD into the intrinsic modal components(IMF) of different characteristic scales and a trend term.On this basis,for each component,combined with other related sequence to construct a Multi-Modal-LSTM model,and finally the predicted value of ammonia nitrogen was obtained through integration.Taking the ammonia nitrogen concentration of two different water quality sections in the Pearl River Delta region as an case for empirical analysis,the experimental results prove that the method has higher model prediction accuracy than the traditional machine learning algorithm and deep learning algorithm,and the prediction effect for the ammonia nitrogen sequence with large data fluctuations and the moment of high concentration is improved significantly,it shows more excellent prediction stability.
作者
张浩彬
薛丽丹
陈光慧
ZHANG Hao-bin;XUE Li-dan;CHEN Guang-hui(School of Economics,Jinan University,Guangzhou 510632,China;Guangdong CreateEnvironmental Technology Company Limited,Guangzhou 510640,China)
出处
《数理统计与管理》
CSSCI
北大核心
2022年第5期761-774,共14页
Journal of Applied Statistics and Management
基金
国家社会科学基金资助项目(18BTJ005)。