摘要
大气污染已经严重影响到人们的生活和健康,大气治理势在必行,探究大气污染物浓度变化的规律,实现污染物浓度预测,对指导大气治理工作具有重要意义。文中构建了一种基于长短期记忆神经网络(Long Short-Term Memory,LSTM)和全连接神经网络(Full Connected,FC)的混合神经网络模型,并提出了数据桶划分的训练方式来解决由于训练数据与预测数据存在较长时间间隔导致精度下降的问题,进而实现大气污染物浓度的预测。该模型具有较好的通用性和精度,充分结合了长短期记忆神经网络和全连接神经网络的优点,能够在多种污染物数据上实现精确预测。以天津市2013-2019年大气污染物数据实现模型的训练和预测,结果表明,混合神经网络模型在PM_(2.5),PM_(10),NO_(2),SO_(2),O_(3),CO 6种污染物浓度的预测上均可以达到R2>0.90,平均百分误差小于15%的效果,LSTM-FC模型在大气污染物预测中具有明显的优势,具有较高的实用价值。
Atmospheric contamination can pose a severe threat to the health of people and incur kinds of diseases,thus,forecasting the concentration of atmospheric contaminant can be of great significance for instructing the atmospheric pollution control.To solve the issue,we propose a kind of mixed forecasting model based on LSTM and full connected neural network,and we introduce the training strategyof data bucket,which can address the issue that the long interval between training data and forecasting sample.Our model has a high performance on both versatility and precision,we fully combine the advantages of LSTM and full connected together and achieve high precision forecasting with varieties of contaminants.Finally,we take an example of forecasting of Tianjin to validate its strength and the results show that our model can achieve R^(2)>0.90,MSE<0.15 performance for all six kinds of pollutant.It shows that LSTM-FC Model has its great strength for atmospheric contaminant concentrations task.
作者
刘梦炀
武利娟
梁慧
段旭磊
刘尚卿
高一波
LIU Meng-yang;WU Li-juan;LIANG Hui;DUAN Xu-lei;LIU Shang-qing;GAO Yi-bo(Tianjin Intelligent Tech Institute of CASIA Tianjin,Tianjin 300300,China;School of Computer Engineering and Science,Shanghai University,Shanghai 200444,China;Institute of Automation,Chinese Academy of Sciences,Beijing 100190,China)
出处
《计算机科学》
CSCD
北大核心
2021年第S01期184-189,共6页
Computer Science
基金
互联网跨界融合创新科技重大专项
大气污染物监测大数据分析平台(18ZXRHSF00250)。
关键词
混合神经网络模型
长短期记忆神经网络
全连接神经网络
污染物浓度预测
多维度特征融合
Hybrid neural network model
Long short-term memory neural network
Full connected neural network
Atmospheric contaminant forecasting
Multi-dimension feature fusion