[1]魏 思,李欣泽*,郤丽媛,等.上下文特征注入融合的空气污染物浓度预测[J].计算机技术与发展,2023,33(09):196-201.[doi:10. 3969 / j. issn. 1673-629X. 2023. 09. 029]
　WEI Si,LI Xin-ze *,XI Li-yuan,et al.Air Pollutant Concentration Prediction Based on Context Feature Injection and Fusion[J].,2023,33(09):196-201.[doi:10. 3969 / j. issn. 1673-629X. 2023. 09. 029]

点击复制

上下文特征注入融合的空气污染物浓度预测()

分享到：

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

33

期数:

2023年09期

页码:

196-201

栏目:

新型计算应用系统

出版日期:

2023-09-10

文章信息/Info

Title:

Air Pollutant Concentration Prediction Based on Context Feature Injection and Fusion

文章编号:

1673-629X(2023)09-0196-06

作者:

魏 思¹ ; 李欣泽2* ; 郤丽媛¹ ; 刘紫君¹ ; 董哲为¹

1. 长安大学 信息工程学院,陕西 西安 710064;

2. 陕西学前师范学院 经济与管理学院,陕西 西安 710100

Author(s):

WEI Si1 ; LI Xin-ze2 * ; XI Li-yuan1 ; LIU Zi-jun1 ; DONG Zhe-wei1

1. School of Information Engineering,Changan University,Xi’an 710064,China;

2. School of Economic and Management,Shaanxi Xueqian Normal University,Xi’an 710100,China

关键词:

空气污染物浓度; PM2. 5; 时间序列; 上下文因素; 特征融合; 门控循环单元

Keywords:

air pollutant concentration; PM2. 5; time series; contextual factors; feature fusion; gated recurrent unit

分类号:

TP183

DOI:

10. 3969 / j. issn. 1673-629X. 2023. 09. 029

摘要:

空气质量预测能够预知区域空间内的大气污染物浓度,对污染防治、环境保护和人身健康等具有非常重要的意义。 针对现有空气污染物预测模型未能充分挖掘和利用上

下文因素的影响和作用,提出了一种上下文特征注入的空气污染物预测模型。 首先,通过循环神经网络和深度置信网络分别学习和提取空气污染物浓度数据的时间序列特征和上下文特征。 然后,使用向量融合机制将提取到的上下文特征注入到时间序列特征中,生成新的融合特征。 最后,将新的高阶融合特征送入预测器,对空气污染物浓度

做出准确可靠的预测。 实验选用 2017 年 1 月至 2021 年 7 月共 55 个月的 PM2. 5 污染物浓度数据,并与 LSTM、GRU、BiLSTM 预测模型相比较,结果表明提出的特征注入模型在多种场景下都能够准确地拟合空气污染物浓度的真实值,预测精度优于传统循环神经网络模型,各项评价指标均较好,表现出较强的适应性和准确性。

Abstract:

Air quality prediction can predict the concentration of air pollutants in regional space,which is of great significance to pollutionprevention,environmental protection?

and human health. Aiming at the failure of existing air pollutant prediction models to fully exploitthe influence and function of context factors, an air pollutant concentration time series prediction model based on contextual featuresinjection was proposed. Firstly,the time series features and context features of air pollutant concentration data were learned and extractedthrough recurrent neural networks and deep belief networks, respectively. Then, the vector fusion mechanism was used to inject theextracted context features into the time series features to generate new fusion features. Finally,the new high-level fusion features weresent to the predictor to make an accurate and reliable prediction of the future air pollutant concentration. 55 months of PM2. 5 pollutantconcentration data from January 2017 to July 2021 were selected for the experiment, and compared with LSTM, GRU and BiLSTMprediction models. It is showed that the proposed feature injected model can accurately approximate the true value of the air pollutantconcentration in a variety of scenarios, the prediction accuracy is better than that of traditional cyclic neural network models, and allevaluation indicators are the best,showing strong adaptability and accuracy.

常用功能