基于RNN的空气污染时空预报模型研究  

Aspatio-temporal prediction framework for air pollution based on deep RNN

作　　者：范竣翔 李琦[1,2] 朱亚杰[1] 侯俊雄 冯逍[1] 

机构地区：[1]北京大学遥感与地理信息系统研究所,北京100871 [2]北京大学智慧城市研究中心,北京100871

出　　处：《测绘科学》2017年第7期76-83,120,共9页Science of Surveying and Mapping

基　　金：国家科技支撑计划项目(2012BAC20B06)

摘　　要：针对空气污染物时间序列中包含缺失值以及现有时间序列预报模型缺乏对时序特征状态建模的问题,该文构建了基于缺失值处理算法和RNN(循环神经网络)的时空预报框架。对空气污染物时序数据设计了3种缺失值处理算法(前向递补、均值替代和权重衰减),用缺失标签和缺失时长对缺失值建模,并在此基础上搭建含有全连接层与LSTM层的深度循环神经网络(DRNN)用于时空预报。使用深度全连接神经网络(DFNN)作为DRNN的对照,用京津冀区域的空气质量和气象数据训练模型,并比较不同模型的预测精度。通过实验,比较了3种缺失值处理方法的效果,结果表明,LSTM在空气污染时空序列预测上的表现优于传统的全连接神经网络层,证实了提出的基于深度学习的时空预报框架的有效性。Time series data in practical applications always contain missing values. In order to handle missing values in time series, as well as the lack of considering temporal attributes in current time series prediction models, we develop a temporal-spatio prediction framework based on missing value processing algorithms and RNN（Recurrent Neural Network）. In this paper, three different missing value modelling algorithms are implemented by using missing tag and missing interval to represent missing pattern in time series data. On top of the missing value modelling algorithms, we construct a deep neural network with LSTM layers and fully connected layers to perform prediction tasks. Real-world air quality and meteoro- logical datasets（Jingjinji Area）are used to train different kinds of deep neural networks, in which the deep feed forward neural networks serve as baseline models. Performances of different models are evaluated in order to compare capabilities of three missing value modelling algorithms, as well as prediction accuracy of different neural network architectures. Experiment results show that deep neural with LSTM layers per- form better than those with only fully connected layers, and validate the capability of the suggested tem- poral-spatio prediction framework based on deep learning.

关 键 词：空气污染 缺失值 RNN LSTM 深度学习 

分 类 号：P208[天文地球—地图制图学与地理信息工程]