Predicting Beijing Air Quality Data Based on LSTM Method

International Journal of Trend in Scientific Research and Development
Volume 5 Issue 3, March-April
@ IJTSRD | Unique Paper ID – IJTSRD40000
Predicting Beijing Air Quality Data Based
School of Information,
ABSTRACT
This paper studies the air quality data of Beijing from 2018 to 2020.
basis of the correlation analysis of pollutant concentration, the circular
neural network model based on LSTM algorithm is built to realize the
prediction of AQI of Beijing. The results show that AQI index has a high
correlation with PM2.5 and PM10, but only has a low negative correlation
with O3. The prediction model of recurrent neural network shows high
prediction accuracy. The research in this paper is helpful to promote the
application of recurrent neural network model in air quality data and ti
series data.
KEYWORDS: AQI; LSTM; Python; Keras; Pearson correlation
1. Research background
With the continuous development of economy and urban
scale, Chinese development has entered a new era, and the
people put forward higher requirements for urban air
quality. As the sandstorm in March 2021, the air quality
problem has once again become the focus of Beijing
citizens. The monitoring and prediction of air quality is
great practical significance in order to improve the air
quality and the level of urban environmental construction.
In order to better monitor and predict air quality, the
national environmental protection department began to
use air quality index (AQI) to quantitatively describe air
quality from 2012. AQI[1] is a kind of conceptual index
which simplifies the concentration of several air pollutants
in conventional monitoring into a single form, and
represents the degree of air pollution and air quality
status by classification. It is suitable for representing the
short-term air quality status and change trend of cities.
With the development of data mining, more and more
machine learning models are applied to the prediction of
air quality. Bai Heming[2] used BP neural network to
forecast the AQI index for different seaso
comparing the forecast value and monitoring value of
different seasons, they concluded that the forecast
accuracy of autumn is the highest. Li Jinglu and Zeng
Tian[3] used the principal component analysis method to
study the air quality data of Beijing from 2000 to 2011,
and concluded that the per capita GDP and the output
value of the tertiary industry had the greatest correlation
with air quality. Wang Mingjie and He Jiajia
method of mathematical statistics and typical circul
classification to study the AQI index. The results showed
April 2021 Available Online: www.ijtsrd.com e
40000 | Volume – 5 | Issue – 3 | March-April
Beijing Air Quality Data Based on LSTM Method
Zeng Guojing, Jin Renhao
f Information, Beijing Wuzi University, Beijing, China
This paper studies the air quality data of Beijing from 2018 to 2020. On the
basis of the correlation analysis of pollutant concentration, the circular
neural network model based on LSTM algorithm is built to realize the
prediction of AQI of Beijing. The results show that AQI index has a high
but only has a low negative correlation
with O3. The prediction model of recurrent neural network shows high
prediction accuracy. The research in this paper is helpful to promote the
application of recurrent neural network model in air quality data and time
AQI; LSTM; Python; Keras; Pearson correlation
How to cite this paper
Jin Renhao "Predicting Beijing Air
Quality Data Based on LSTM Method"
Published in
International
Journal of Trend in
Scientific Research
and Development
(ijtsrd), ISSN: 2456
6470, Volume
Issue-3, April 2021,
pp.774-777, URL:
www.ijtsrd.com/papers/ijtsrd40000.pdf
Copyright © 20
International Journal of Trend in
Scientific Research and Development
Journal. This is an Open Access article
distributed under
the terms of the
Creative Commons
Attribution License
(http://creativecommons
With the continuous development of economy and urban
scale, Chinese development has entered a new era, and the
requirements for urban air
quality. As the sandstorm in March 2021, the air quality
problem has once again become the focus of Beijing
citizens. The monitoring and prediction of air quality is
great practical significance in order to improve the air
y and the level of urban environmental construction.
In order to better monitor and predict air quality, the
national environmental protection department began to
use air quality index (AQI) to quantitatively describe air
d of conceptual index
which simplifies the concentration of several air pollutants
in conventional monitoring into a single form, and
represents the degree of air pollution and air quality
status by classification. It is suitable for representing the
term air quality status and change trend of cities.
With the development of data mining, more and more
machine learning models are applied to the prediction of
used BP neural network to
forecast the AQI index for different seasons in Beijing. By
comparing the forecast value and monitoring value of
different seasons, they concluded that the forecast
accuracy of autumn is the highest. Li Jinglu and Zeng
used the principal component analysis method to
ata of Beijing from 2000 to 2011,
and concluded that the per capita GDP and the output
value of the tertiary industry had the greatest correlation
with air quality. Wang Mingjie and He Jiajia[4] used the
method of mathematical statistics and typical circulation
classification to study the AQI index. The results showed
that the main pollutants causing weather pollution were
NO 、PM . and O .Li
built a fractal popular learning support vector machine to
predict AQI index. They adopt t
dimension first and then reduce the dimension, which
improved the accuracy and stability of prediction. Xu Qi
and Wu Qizhong[6] used the comprehensive scoring
method to monitor and forecast the PM
the air. Based on the WRF
evaluation results showed that the accuracy was better
than the official forecast.
However, the air pollution index is a typical time series
data. When using the traditional statistical model and the
common neural network method to predict, the accuracy
is not high enough and the calculation time is long.
Recurrent neural network is a kind of neural network
model with the input of time series data, which is more
suitable for the modeling and prediction of time series
data. LSTM solves the common problems of gradient
disappearance and gradient explosion in traditional
recurrent neural network. It is a common recurrent neural
network algorithm and has many successful
applications[7]-[10] in predicting time series data. But at
present, the research on the application of recurrent
neural network model based on LSTM algorithm in air
quality prediction is still lacking, especially in Beijing data.
Therefore, this paper uses Python deep learning library
keras to build LSTM recurrent
realize the prediction of Beijing air quality data, and
selects AQI as the main index of air quality as the
prediction target variable.
International Journal of Trend in Scientific Research and Development (IJTSRD)
e-ISSN: 2456 – 6470
April 2021 Page 774
n LSTM Method
China
How to cite this paper: Zeng Guojing |
Jin Renhao "Predicting Beijing Air
Data Based on LSTM Method"
Published in
International
Journal of Trend in
Scientific Research
and Development
(ijtsrd), ISSN: 2456-
6470, Volume-5 |
3, April 2021,
777, URL:
www.ijtsrd.com/papers/ijtsrd40000.pdf
Copyright © 2021 by author (s) and
International Journal of Trend in
Scientific Research and Development
This is an Open Access article
distributed under
the terms of the
Creative Commons
Attribution License (CC BY 4.0)
//creativecommons.org/licenses/by/4.0)
that the main pollutants causing weather pollution were
Ping and Ni Zhiwei[5]
built a fractal popular learning support vector machine to
predict AQI index. They adopt the method of fractal
dimension first and then reduce the dimension, which
improved the accuracy and stability of prediction. Xu Qi
used the comprehensive scoring
method to monitor and forecast the PM2.5 concentration in
WRF-CMAQ model system, their
evaluation results showed that the accuracy was better
However, the air pollution index is a typical time series
data. When using the traditional statistical model and the
ethod to predict, the accuracy
is not high enough and the calculation time is long.
Recurrent neural network is a kind of neural network
model with the input of time series data, which is more
suitable for the modeling and prediction of time series
STM solves the common problems of gradient
disappearance and gradient explosion in traditional
recurrent neural network. It is a common recurrent neural
network algorithm and has many successful
in predicting time series data. But at
resent, the research on the application of recurrent
neural network model based on LSTM algorithm in air
quality prediction is still lacking, especially in Beijing data.
Therefore, this paper uses Python deep learning library
keras to build LSTM recurrent neural network model to
realize the prediction of Beijing air quality data, and
selects AQI as the main index of air quality as the
IJTSRD40000

2. Theoretical basis
2.1. Keras
Keras is a powerful high-level neural network API written
for python. It can use tensor flow, theano and cntk as the
interfaces of high-level applications. Keras is one of the
commonly used machine learning tools, which has four
advantages: user-friendly, modular operation, strong
scalability, and high collaboration with Pyt
a large number of functions and program optimizers and
other components. The optimizer included in Keras can
realize back propagation algorithm and adaptive gradient
descent algorithm, which is convenient for the
implementation of LSTM recurrent neural network
algorithm.
2.2. Principle of LSTM neural network
Long term and short-term memory network (LSTM) is a
variant algorithm of recurrent neural network (RNN). By
using time back propagation training, it can solve the
problems of gradient disappearance and gradient
explosion in common neural network method. It is widely
used in image video recognition, stock price trend
prediction, disease prediction and other fields. LSTM
algorithm uses memory cells to replace conventional
neurons in RNN. Memory cells are more flexible
components than neurons, and memory modules are
introduced. Each storage unit is composed of forgetting
Fig 2 Variation trend of AQI index and six pollutants
International Journal of Trend in Scientific Research and Development (IJTSRD) @ www.ijtsrd.com
level neural network API written
It can use tensor flow, theano and cntk as the
level applications. Keras is one of the
commonly used machine learning tools, which has four
friendly, modular operation, strong
scalability, and high collaboration with Python. It contains
a large number of functions and program optimizers and
other components. The optimizer included in Keras can
realize back propagation algorithm and adaptive gradient
descent algorithm, which is convenient for the
urrent neural network
Principle of LSTM neural network
term memory network (LSTM) is a
neural network (RNN). By
using time back propagation training, it can solve the
earance and gradient
explosion in common neural network method. It is widely
used in image video recognition, stock price trend
prediction, disease prediction and other fields. LSTM
algorithm uses memory cells to replace conventional
cells are more flexible
components than neurons, and memory modules are
introduced. Each storage unit is composed of forgetting
gate, input gate and output gate, and its structure is shown
in Figure 1.In Fig 1:t represents the specific time,
and represent the input sequence at t time,
time and 1 time respectively;
represent the outputs of the memory cells at t time
time and t 1time respectively. The
hyperbolic tangent function and
activation function. This function can transform to
produce a smooth range value between 0 and 1, so as to
observe the change of output value when the input value
changes slightly.
3. Construction of LSTM prediction model
3.1. data sources
This paper is based on the air quality data of Beijing from
January 2018 to December 2020, and the data is from the
website of China Weather Post
(http://www.tianqihoubao.com/).A total of 1096 rows of
observations were obtained.
daily AQI index and concentrations of six pollutants
NO 、PM . 、SO 、O 、PM
sampling time and force majeure and other factors, some
date data are missing. This paper uses the monthly mean
of these seven kinds of data to borrow and supple
missing values. The trend of AQI index and six kinds of
pollutant values is shown in Figure 2.
Fig 1 The structure of LSTM
Variation trend of AQI index and six pollutants
www.ijtsrd.com eISSN: 2456-6470
April 2021 Page 775
gate, input gate and output gate, and its structure is shown
in Figure 1.In Fig 1:t represents the specific time, 、
represent the input sequence at t time, t 1
time respectively;h 、h and h
represent the outputs of the memory cells at t time、t 1
time respectively. The is the
hyperbolic tangent function and is the sigmoid
activation function. This function can transform to
produce a smooth range value between 0 and 1, so as to
observe the change of output value when the input value
Construction of LSTM prediction model
er is based on the air quality data of Beijing from
January 2018 to December 2020, and the data is from the
website of China Weather Post
(http://www.tianqihoubao.com/).A total of 1096 rows of
observations were obtained. Data information includes
index and concentrations of six pollutants CO、
in Beijing.Due to the long
sampling time and force majeure and other factors, some
date data are missing. This paper uses the monthly mean
of these seven kinds of data to borrow and supplement the
missing values. The trend of AQI index and six kinds of
pollutant values is shown in Figure 2.

3.2. Correlation analysis between AQI
It can be seen from Fig 2 that the change trend of AQI and the concentrations of
is roughly the same, When the AQI index becomes higher, the other five pollutants will also become higher. When the AQI
index becomes lower, the other five pollutants will also become lower.
AQI index and the concentrations of CO、
concentration of O becomes lower, so there is a negative correlation between AQI index and
further analyze the relationship between AQI and
of each index is shown in Table 1.There was
NO 、PM . 、SO andPM , and a weak negative correlation between AQI index and
value of - 0.08.PM2.5 and PM10 had the highest positive correlation with
were 0.936 and 0.785.Therefore, in the study of air pollution control in Beijing, we can formulate relevant policies from the
perspective of controlling the emission of these two pollutants, and take certai
these two pollutants.
Table1. Correlation coefficient matrix of AQI index and six pollutants in Beijing
AQI
AQI 1
PM . 0.936
SO 0.438
NO 0.580
PM 0.785
O -0.080
CO 0.757
4. Research on AQI prediction
According to the correlation analysis of AQI index and six kinds of common pollutants, the air quality of the next day can
be predicted by the historical data of these pollutant concentration indexes.
is the main index to measure air quality. The next day's AQI index value is used as the prediction target variable, and the
AQI index and the historical index value of six pollutants are used as the model input variables. The LSTM neural
algorithm program is supported by using Keras module in Python.
value, this paper uses the method of maximum and minimum to realize the normalization of each index data.
model, there are 100 neurons in the hidden layer and only one neuron in the output layer; the first 70% of the sample data
is used as training data, and the last 30% as test data.
results of the model and the real values, the predicted results are de normalized.
value and the real value on the training set and the test set is shown in Figure 3. It can be seen from the figure that the
prediction error of LSTM model on the training set and the test set is small, indicating that the model has high prediction
accuracy. The average absolute error of the model in the training set and the test set are 3.31 and 5.17 respectively, and
the average absolute error rate in the train
that the model has high prediction accuracy.
Fig 3 Prediction effect
International Journal of Trend in Scientific Research and Development (IJTSRD) @ www.ijtsrd.com
Correlation analysis between AQI index and pollutants
It can be seen from Fig 2 that the change trend of AQI and the concentrations of CO、NO 、
When the AQI index becomes higher, the other five pollutants will also become higher. When the AQI
index becomes lower, the other five pollutants will also become lower. Therefore, there is a positive correlation between
、NO 、PM . 、SO andPM .However, when the AQI index becomes higher, the
becomes lower, so there is a negative correlation between AQI index and
further analyze the relationship between AQI and CO、NO 、PM . 、SO 、O 、PM , the Pearson correlation coefficient
of each index is shown in Table 1.There was a positive correlation between AQI index and the concentrations of
, and a weak negative correlation between AQI index and O concentration, with the coefficient
0.08.PM2.5 and PM10 had the highest positive correlation with AQI, and the correlation coefficients respective
perspective of controlling the emission of these two pollutants, and take certain measures to reduce the concentration of
. ! "# $ CO
0.936 0.438 0.580 0.785 -0.080 0.757
1 0.492 0.659 0.624 -0.043 0.857
0.492 1 0.619 0.413 -0.258 0.624
0.659 0.619 1 0.503 -0.453 0.718
0.624 0.413 0.503 1 -0.003 0.474
-0.043 -0.258 -0.453 -0.003 1 0.474
0.857 0.624 0.718 0.464 -0.172 1
be predicted by the historical data of these pollutant concentration indexes. This paper establishes a mod
AQI index and the historical index value of six pollutants are used as the model input variables. The LSTM neural
algorithm program is supported by using Keras module in Python. Due to the difference of data scale between each index
value, this paper uses the method of maximum and minimum to realize the normalization of each index data.
e are 100 neurons in the hidden layer and only one neuron in the output layer; the first 70% of the sample data
is used as training data, and the last 30% as test data. Finally, when comparing the difference between the predicted
he real values, the predicted results are de normalized. The fitting curve between the predicted
he training set and the test set is small, indicating that the model has high prediction
The average absolute error of the model in the training set and the test set are 3.31 and 5.17 respectively, and
the average absolute error rate in the training set and the test set are 4.13% and 4.91% respectively, which further shows
that the model has high prediction accuracy. In Figure 3, green represents the training set and red represents the test set
Prediction effect of LSTM model on training set and test set
www.ijtsrd.com eISSN: 2456-6470
April 2021 Page 776
、PM . 、SO andPM in Beijing
When the AQI index becomes higher, the other five pollutants will also become higher. When the AQI
Therefore, there is a positive correlation between
However, when the AQI index becomes higher, the
becomes lower, so there is a negative correlation between AQI index and O concentration.In order to
the Pearson correlation coefficient
a positive correlation between AQI index and the concentrations of CO、
concentration, with the coefficient
AQI, and the correlation coefficients respective
n measures to reduce the concentration of
CO
0.757
0.857
0.624
0.718
0.474
0.474
1
This paper establishes a model for AQI, which
AQI index and the historical index value of six pollutants are used as the model input variables. The LSTM neural network
Due to the difference of data scale between each index
value, this paper uses the method of maximum and minimum to realize the normalization of each index data. In the LSTM
e are 100 neurons in the hidden layer and only one neuron in the output layer; the first 70% of the sample data
Finally, when comparing the difference between the predicted
The fitting curve between the predicted
he training set and the test set is small, indicating that the model has high prediction
The average absolute error of the model in the training set and the test set are 3.31 and 5.17 respectively, and
ing set and the test set are 4.13% and 4.91% respectively, which further shows
In Figure 3, green represents the training set and red represents the test set
of LSTM model on training set and test set

International Journal of Trend in Scientific Research and Development (IJTSRD) @ www.ijtsrd.com eISSN: 2456-6470
@ IJTSRD | Unique Paper ID – IJTSRD40000 | Volume – 5 | Issue – 3 | March-April 2021 Page 777
5. Conclusion
Based on the analysis of the concentration of air pollutants
in Beijing from January 2018 to December 2020, this
paper analyzes the air pollution index, the concentration
change trend of six pollutants and the correlation of each
pollutant index. The results show that there is a positive
correlation between AQI and the concentrations of CO、
NO 、PM . 、SO 、O 、PM , and a negative correlation
between AQI and O .Due to the nonlinear relationship
between the AQI index and the concentration of these
pollutants, the traditional statistical prediction method
cannot achieve the ideal prediction accuracy. In this paper,
the recurrent neural network model is used to establish
the prediction model, and the long-term and short-term
memory network (LSTM) is used for model operation. The
results show that the model has high prediction accuracy.
The results show that the recurrent neural network can be
widely used in the area of air quality data prediction, and
can also be extended to more time series data.
Acknowledgment
This paper was supported in part by General Project of
Science and Technology Plan of Beijing Municipal
Education Commission (KM201910037002), and Beijing
Excellent Talents Training Funding
(2017000020124G051).
Reference
[1] Liu-Jie, Yang-Peng, Lu Wen-sheng, et al.
Environmental air quality evaluation method based
on the six pollutants in the urban areas of Beijing
[J]. Journal of Safety and Environment, 2015, 15(1):
310-315
[2] BAI Heming, SHEN Runping, SHI Huading, et al.
Forecasting model of air pollution index based on
BP neural network[J]. Environmental Science &
Technology，2013, 36(3): 186-189
[3] LI Jinglu, ZENG Tian. Analysis on the Principal
Component of Factors Affecting Air Quality in
Beijing: From 2000-2011 Years of Experience
Data[J]. Ecological Economy, 2017, 33(1): 169-173
[4] WANG Mingjie, HE Jiajia, WANG Shuxin, ZHANG Lei.
2018. Atmospheric pollution characteristics and
typical circulation pattern in Shenzhen based on
AQI [J]. Ecology and Environmental Sciences, 27(2):
268-275.
[5] LI Ping, NI Zhiwei, ZHU Xuhui, WU Zhangjun. Air
pollution index prediction model of SVM based on
fractal manifold learning[J]. Journal of Systems
Science and Mathematical Sciences, 2018, 38(11):
1296-1306.
[6] XU Qi, WU Qizhong, LI Dongqing, et al. 2020.
Assessment of the Air Quality Numerical Forecast in
the Main District of Beijing (2018) [J]. Climatic and
Environmental Research (in Chinese), 25 (6):
616−624.
[7] ZHANG Zhen, ZHU Quanjie, LI Qingsong, et al.
Prediction of mine gas concentration in heading
face based on keras long short time memory
network[J]. Safety and Environmental Engineering,
2021, 20(1): 61-67
[8] Yang Taichun，Tao Jianfeng，Yu Honggan，Liu
Chengliang. Real-time prediction of torque of cutter
head of shield machine based on LSTM[J]. 2020,
16（6）: 1801-1808.
[9] Zhiling Tang, Qianqian Liu, Minjie Wu, Wenjing
Chen, Jingwen Huang. WiFi CSI Gesture Recognition
Based on Parallel LSTM-FCN Deep Space-Time
Neural Network[J]. China Communications, 2021,
18(03): 205-215.
[10] ZHANGLin, HUANG Yanwen, XUAN Jie, FU Xiong,
LIN Qiaomin, WANG Ruchuan. Trust Evaluation
Model Based on PSO and LSTM for Huge
Information Environments[J]. Chinese Journal of
Electronics, 2021, 30(01): 92-101.

Predicting Beijing Air Quality Data Based on LSTM Method

More Related Content

Similar to Predicting Beijing Air Quality Data Based on LSTM Method (20)

More from YogeshIJTSRD (20)

Recently uploaded (20)

Predicting Beijing Air Quality Data Based on LSTM Method