globalchange  > 气候变化与战略
DOI: 10.1016/j.atmosenv.2019.117072
A novel framework for daily forecasting of ozone mass concentrations based on cycle reservoir with regular jumps neural networks
Author: Mo Y.; Li Q.; Karimian H.; Fang S.; Tang B.; Chen G.; Sachdeva S.
Source Publication: Atmospheric Environment
ISSN: 1352-2310
Publishing Year: 2020
Volume: 220
Language: 英语
Keyword: Air pollution ; Benchmarking ; Deep neural networks ; Learning systems ; Machine learning ; Ozone ; Recurrent neural networks ; Time series ; Adverse health effects ; Air pollution predictions ; Classification rates ; Concentration prediction ; Machine learning methods ; Ozone concentration ; Surface concentration ; Time series features ; Forecasting ; ozone ; artificial neural network ; atmospheric pollution ; concentration (composition) ; machine learning ; ozone ; prediction ; air monitoring ; air pollution ; air quality standard ; Article ; China ; decomposition ; forecasting ; intrinsic mode function ; machine learning ; priority journal ; recurrent neural network ; support vector machine ; time series analysis ; Beijing [China] ; China
Subject Area: Air pollution ; Deep neural networks ; Machine learning ; Ozone ; Prediction ; Recurrent neural networks
Abstract: An understanding of the growth in surface concentration of ozone and its adverse health effects are important for environmental departments to make sensible decisions for future. Our hybrid model CEEMD+CRJ+MLR is the first attempt to improve CRJ in the field of air pollution prediction and ozone forecasting. For this novel framework, CEEMD has been adopted to decompose original MDA8_O3 history into several sub-series. After that, for each IMF, CRJ is used to extract time-series features. These time-series features are fed into appropriate machine learning methods for prediction. In addition to that, residual is also predicted through normal methods. A model, which is trained with data from 1 May 2014 to 31 May 2017, is validated with data from 1 June 2017 to 30 May 2018, obtained from four stations of Beijing, China. The hybrid model has input variables which are combined with related pollutants, meteorological forecasts and UV index, and predict maximum daily 8-h average ozone (MDA8_O3) concentration in different time intervals. Our experimental results show that the CEEMD+CRJ+MLR model exhibits the best performance compared with other benchmark models generally. For four stations, IA, MAE, RMSE and MAPE average of +1 (forecasting 1 day in advance) are 0.9763, 12.84, 17.81 and 18.5% respectively and of +2 are 0.9679, 15.17, 20.15 and 23.86% respectively. Especially in the case of forecasting heavy ozone concentration (Level III), a critical issue in air pollution predictions, the classification rate of our hybrid model has improved from 29.4% (for CRJ) to 83.4% in +1 and from 38% (for CRJ) to 73% in +2. For long time forecasting, the CEEMD+CRJ+MLR also shows its outstanding performance in whole levels and level III ozone concentration. Our hybrid model, with accurate and stable results, is highly effective for MDA8_O3 concentration prediction and can efficiently be applied in other regions. © 2019 Elsevier Ltd
Citation statistics:
Document Type: 期刊论文
Appears in Collections:气候变化与战略

Files in This Item:

There are no files associated with this item.

Affiliation: Institute of Remote Sensing and Geographical Information Systems, School of Earth and Space Sciences, Peking University, Beijing, China; School of Architecture, Surveying and Mapping Engineering, Jiangxi University of Science and TechnologyJiangxi, China; Shenzhen Engineering Laboratory of Ocean Environmental Big Data Analysis and Application, Shenzhen, 518055, China

Recommended Citation:
Mo Y.,Li Q.,Karimian H.,et al. A novel framework for daily forecasting of ozone mass concentrations based on cycle reservoir with regular jumps neural networks[J]. Atmospheric Environment,2020-01-01,220
Recommend this item
Sava as my favorate item
Show this item's statistics
Export Endnote File
Google Scholar
Similar articles in Google Scholar
[Mo Y.]'s Articles
[Li Q.]'s Articles
[Karimian H.]'s Articles
Similar articles in Baidu Scholar
[Mo Y.]'s Articles
[Li Q.]'s Articles
[Karimian H.]'s Articles
CSDL cross search
Similar articles in CSDL Cross Search
[Mo Y.]‘s Articles
[Li Q.]‘s Articles
[Karimian H.]‘s Articles
Related Copyright Policies
所有评论 (0)

Items in IR are protected by copyright, with all rights reserved, unless otherwise indicated.