电子学报 ›› 2022, Vol. 50 ›› Issue (6): 1410-1427.DOI: 10.12263/DZXB.20200258

所属专题: 长摘要论文

• 学术论文 • 上一篇    下一篇

一种卷积自编码深度学习的空气污染多站点联合预测模型

张波1,3, 陆云杰1, 秦东明2,4, 邹国建1   

  1. 1.上海师范大学信息与机电工程学院,上海 200234
    2.同济大学电子与信息工程学院,上海 201804
    3.上海智能教育大数据工程技术研究中心,上海 200234
    4.中科三清科技有限公司,北京 100089
  • 收稿日期:2020-03-14 修回日期:2021-07-25 出版日期:2022-06-25
    • 作者简介:
    • 张 波 男,1979年11月出生,江苏常州人.同济大学计算机应用技术专业工学博士.现为上海师范大学信息与机电工程学院院长、教授.主要研究方向为新一代人工智能与深度学习技术等.E-mail: zhangbo@shnu.edu.cn
      陆云杰 男,1995年6月出生,江苏苏州人.上海师范大学信息与机电工程学院计算机应用技术工学硕士.主要研究方向为机器学习.E-mail: 1000459475@smail.shnu.edu.cn
      秦东明 男,1981年4月出生,河南社旗人.同济大学电子与信息工程学院工学博士.主要研究方向为高性能计算、人工智能、深度学习等.
      邹国建 男,1993年11月出生,安徽滁州人.上海师范大学信息与机电工程学院计算机应用技术专业工学硕士.主要研究方向为深度学习、知识图谱.
    • 基金资助:
    • 国家自然科学基金 (61802258); 上海自然科学基金 (18ZR1428300)

A Multi-Site Joint Air Pollution Prediction Model Based on Convolutional Auto-Encoder Deep Learning

ZHANG Bo1,3, LU Yun-jie1, QIN Dong-ming2,4, ZOU Guo-jian1   

  1. 1.College of Information, Mechanical and Electrical Engineering, Shanghai Normal University, Shanghai 200234, China
    2.College of Electronic and Information Engineering, Tongji University, Shanghai 201804, China
    3.Shanghai Engineering Research Center of Intelligent Education and Bigdata, Shanghai Normal University, Shanghai 200234, China
    4.3Clear Science & Technology Co. Ltd, Beijing 100193, China
  • Received:2020-03-14 Revised:2021-07-25 Online:2022-06-25 Published:2022-06-25
    • Supported by:
    • National Natural Science Foundation of China (61802258); Natural Science Foundation of Shanghai Municipality, China (18ZR1428300)

摘要:

城市空气污染因空间扩散特性呈现出区域内的浓度高关联性.因此如何通过多个空气污染监测站的时空数据预测特定目标地点的污染情况,以解决站点分布不匀的问题,是一个重要的研究工作.本文结合空气污染物因素特性和气象因素的多维度影响,提出了一个利用区域内多站点空间监测数据实现特定目标站点的空气污染物浓度预测模型.该模型通过多层卷积神经网络(Convolutional Neural Network,CNN)实现城市多站点污染物浓度与气象数据之间的维度关联特征及空间关联特征学习,进而利用基于多层长短期记忆网络(Long Short-Term Memory,LSTM)的自编码网络实现多站点浓度的时序关联特征分析.实验通过真实数据集验证,所提出的预测模型获得了高于传统机器学习污染物浓度预测模型的预测准确度,且在多个城市数据集上验证了模型的泛化能力.

长摘要
随着空气污染问题日益严重,如何利用相关数据实现空气污染高效分析和预测,已成为实现空气污染高效治理所必须解决的迫切问题。城市空气污染因空间扩散特性呈现出区域内的浓度高关联性,而之前的研究大多聚焦于以单城市内的综合污染物数据来进行单城市的污染物浓度预测,并未考虑城市内多个分布不均匀站点之间的时空关系特征并进行联合预测。因此本文针对此瓶颈问题,提出了一个利用区域内多站点空间监测数据实现特定目标站点的空气污染物浓度预测模型CAELearning。该模型通过多层卷积神经网络(CNN)实现城市多站点污染物浓度与气象数据之间的维度关联特征及空间关联特征学习,进而利用基于多层长短期记忆网络(LSTM)的端到端模型架构网络实现多站点浓度的时序关联特征分析。模型同时考虑到了空间关联性即城市内多站点之间污染物浓度的相互影响,和时间关联性即污染物浓度前后时间段内的相互影响。通过真实数据集验证,该模型可以对未来一段时间内污染物浓度进行连续性预测,获得了高于传统机器学习污染物浓度预测模型的预测准确度,并在不同城市的预测上具有较好的泛化能力。同时,模型已经多次在国家级区域空气污染监测预报任务中作为实际辅助模型之一得到应用,体现出较好的价值。

关键词: 深度学习, 空气污染, 时空数据, 多站点联合预测, 卷积神经网络, 长短期记忆网络

Abstract:

Due to the nature of spatial diffusion, urban air pollution presents a high correlation with regional concentration feature. Therefore, how to use spatio-temporal related data from urban multiple air pollution detection sites to predict air pollution concentration of a special target location is an important research effort for solving the problem of uneven site distribution. Concerning the multi-dimensional impacts of the air pollutant factors’ features and the influence of meteorological factors, we propose an air pollutant concentration prediction model, which uses multi-site spatial detection data within a region to predict the concentration of the target station in this paper. This model is able to learn dimensional correlation characteristics and spatial correlation characteristics from multi-site pollutant concentration and meteorological data in the urban area through the multi-layer convolutional neural network, and then analyze the time-series correlation characteristics of multi-site concentration by utilizing the multi-layer auto-encoder network based long short-term memory network. The experimental results show that our proposed model obtains better performances than traditional machine learning based models under a real-world dataset, and meanwhile, the generalization performance of the proposed model has been examined based on multiple urban air pollution database.

Extended Abstract
As the air pollution problem becomes more and more serious, how to use relevant data to achieve efficient air pollution analysis and prediction has become an urgent problem that must be solved to achieve effective air pollution management. Urban air pollution exhibits high intra-regional concentration correlations due to spatial dispersion characteristics, while most previous studies focus on single-city pollutant concentration prediction with integrated pollutant data within a single city, without considering the characteristics of spatio-temporal relationships between multiple unevenly distributed stations within a city and making joint predictions. Therefore, this paper addresses this bottleneck problem and proposes a model called CAELearning (Learning net based on CNN and Auto-Encoder) for air pollutant concentration prediction at specific target sites using multi-site spatial monitoring data in the region by combining the characteristics of air pollutant factors and the multidimensional influence of meteorological factors. The proposed model considers the coupling relationship between the spatio-temporal characteristics of pollutants and meteorological data from the serial perspective of the model, realizes the dimensional association features and spatial association features learning between urban multi-site pollutant concentrations and meteorological data through a multilayer convolutional neural network (CNN), and then uses an end-to-end model architecture network based on a multilayer long short-term memory network (LSTM) to realize the temporal association features analysis of multi-site concentrations. The model takes into account both spatial correlation, i.e., the interaction of pollutant concentrations among multiple stations within a city, and temporal correlation, i.e., the interaction of pollutant concentrations in the time period before and after. Experiments based on real datasets indicate that the proposed prediction model can predict pollutant concentrations at specific sites in the next 24 hours based on pollutant concentrations at multiple sites in the past 72 hours as well as meteorological data, obtaining higher prediction accuracy than traditional machine learning pollutant concentration prediction models, and validating the generalization capability on multiple urban datasets. Meanwhile, the model has been applied as one of the practical auxiliary models in the national-level regional air pollution monitoring and forecasting tasks for several times, reflecting a better value.

Key words: deep learning, air pollution, spatio-temporal data, multiple sites joint prediction, convolutional neural network, long short-term memory network

中图分类号: