Professional Documents
Culture Documents
Modeling Air Quality Prediction Using A Deep Learning Approach Method Optimization and Evaluation
Modeling Air Quality Prediction Using A Deep Learning Approach Method Optimization and Evaluation
Wenjing Mao, Weilin Wang, Limin Jiao, Suli Zhao, Anbao Liu
PII: S2210-6707(20)30785-X
DOI: https://doi.org/10.1016/j.scs.2020.102567
Reference: SCS 102567
This is a PDF file of an article that has undergone enhancements after acceptance, such as
the addition of a cover page and metadata, and formatting for readability, but it is not yet the
definitive version of record. This version will undergo additional copyediting, typesetting and
review before it is published in its final form, but we are providing this version to give early
visibility of the article. Please note that, during the production process, errors may be
discovered which could affect the content, and all legal disclaimers that apply to the journal
pertain.
Wenjing Maoa, b,1, Weilin Wanga, b,1, Limin Jiaoa, b,*, Suli Zhaoa, Anbao Liuc
China
of
University, Wuhan, 430079, China
ro
* Corresponding author
-p
E-mail addresses: Wenjing Mao [email protected]; Weilin Wang
re
[email protected]; Limin Jiao [email protected]; Suli Zhao [email protected];
Competing Interests
1 These authors contributed equally to this work and should be considered as co-first
ur
authors.
Jo
Highlights
long-term predictions.
time series.
Abstract: Air pollution is one of the hot issues that attracted widespread attention from
of
urban and society management. Air quality prediction is to issue an alarm when severe
pollution occurs, or pollution concentration exceeds a specific limit, contributing to the
ro
measure-taking of relevant departments, guiding urban socio-economic activities to
promote sustainable urban development. However, existing methods have failed to
make full use of the temporal features from spatiotemporal correlations of air quality
-p
monitoring stations, and achieved poor performances in long-term predictions (up to or
above 24h-predictions). In this study, we proposed a deep learning framework to predict
air quality in the following 24 hours: a neural network with a temporal sliding long
re
short-term memory extended model (TS-LSTME). The model integrated the optimal
time lag to realize sliding prediction through multi-layer bidirectional long short-term
memory (LSTM), involving the hourly historical PM2.5 concentration, meteorological
lP
data, and temporal data. We applied the proposed model to predict the next 24-hour
average PM2.5 concentration in Jing-Jin-Ji region, with the most severe air pollution in
China. The proposed model had better stability and performances with high correlation
coefficient R2 (0.87), compared to the multiple linear regression (MLR), the support
na
vector regression (SVR), the traditional LSTM, and the long short-term memory
extended (LSTME) models. Moreover, the proposed model can achieve PM2.5
concentration predictions with high accuracy in long-term series (48 h and 72 h). We
ur
also tested the model to predict O3 concentration. The proposed model could be applied
for other air pollutants. The proposed methods can significantly improve air quality
prediction information services for the public and provide support for early warning
Jo
Keywords: air pollutant; air quality prediction; deep learning; long-term prediction;
temporal sliding
1. Introduction
With the rapid development of the world economy and the acceleration of
industrialization and urbanization, many cities around the world are suffering from air
pollution. Air pollution problems are becoming more apparent, threatening human
production, life, and sustainable social development seriously (Li et al., 2019). Wherein,
PM2.5 (fine particles with a diameter of 2.5 microns or less in the atmosphere) is the
main factor of air pollution, and the increasing PM2.5 concentration will also directly
affect the human health (Kampa and Castanas, 2008; Martins and Carrilho da Graça,
2018; Kumar et al., 2020; Barzeghar et al., 2020; Pak et al., 2020). Therefore, real-time
PM2.5 concentration predictions in advance have significant practical and social values,
which play a significant role in making atmospheric management decisions to control
air pollution and are important information needed to facilitate sustainable development.
Previous models for air pollutant predictions can be divided into two major
categories: physical prediction models and statistical prediction models. The physical
prediction models are based on aerodynamics, atmospheric physics, and chemistry to
study pollutant diffusion mechanism (Geng et al., 2015), and apply mathematical
methods to calculate the spatiotemporal distributions of pollutants (Jeong et al., 2011;
Kim et al., 2010), such as CMAQ models (Jiang and Yoo, 2018; Woody et al., 2016;
of
Yang et al., 2019), WRF-Chem models (Geng et al., 2015; Liu et al., 2018; Ma et al.,
2018) and GEO-Chem models (Jeong et al., 2011; Lee et al., 2017; Wang et al., 2014).
ro
However, due to the complex pollutant diffusion mechanism, some problems produced
by these models result in the usage limitations, such as colossal calculation workload,
the complexity of the process, uncertainty of parameters (Pan et al., 2011; Stern et al.,
-p
2008; Vautard et al., 2007; J. Wang et al., 2019). The statistical prediction models are
based on statistics and use historical time series data to predict future air quality.
Compared with the physical prediction models, the statistical prediction models not
re
only avoid the complicated mechanism so that the calculation is faster and the cost of
each prediction can be reduced, but also can achieve an almost equivalent level of PM2.5
concentration prediction accuracy of the physical prediction models (Engel-Cox et al.,
lP
2013). Due to those advantages, the application of the statistical prediction models is
more extensive. Commonly used statistical prediction models are included machine
learning (ML) and artificial neural network (ANN) models (Perez and Reyes, 2006).
Machine learning models include multiple linear regression (MLR) models (Han et al.,
na
2020; Stadlober et al., 2008), support vector regression (SVR) models (Zhu et al., 2018),
random forest models (Gariazzo et al., 2020), etc. Machine learning models have clear
mathematical logic, in which the relationship between input and output is relatively
ur
definite, and the structures are simple. However, these models ignore the influences
of external variables and spatial factors (Leng et al., 2017), and are prone to overfitting
due to improper index selection (Ausati and Amanollahi, 2016).
Jo
of
models are based on the LSTM, considering the meteorological factors and the
spatiotemporal correlations with the monitoring stations, and they performed well in
ro
short-term prediction tasks (Li et al., 2017). In recent years, some researchers combined
convolutional neural networks (CNN) and LSTM models to simulate the changing
process of air pollutants and achieved good performances of prediction, such as C-
-p
LSTME (Wen et al., 2019), graph GC-LSTM (Qi et al., 2019), multi-output deep
learning LSTM (DM-LSTM) (Zhou et al., 2019), and Deep CNN-LSTM (Huang and
Kuo, 2018) models. However, they have failed to capture the time dependence features
re
effectively, which resulted in exhibiting relatively low precision for long-term
prediction tasks (up to or above 24h-predictions) of PM2.5 concentration.
Studies have proven that spatiotemporal correlations of PM2.5 monitoring stations
lP
play a non-substitutable role, which can significantly affect the changing trend of PM2.5
concentration (Li et al., 2017; W. Wang et al., 2019; Wen et al., 2019). In our study, we
proposed a temporal sliding long short-term extended (TS-LSTME) model, taking the
spatiotemporal correlations of PM2.5 concentration at the monitoring stations into
na
integrated to improve the prediction performance of the model. The model was applied
to the long-term predictions of PM2.5 concentration in Jing-Jin-Ji (Beijing, Tianjin, and
Hebei) area, China. The results showed that the proposed model with high accuracy
Jo
could be used for long-term predictions of PM2.5 concentration. Compared with the
MLR model, SVR model, LSTME model, and conventional LSTM model, the TS-
LSTME model achieved the best performance, especially in the predictions of long-
term series (48 hours and 72 hours). According to the seasonal characteristics of PM2.5
concentration, we selected representative time points (January 1st, April 1st, July 1st,
and October 1st) to spatially interpolate the model predicted results and observed values
of monitoring stations and compared spatiotemporal distributions. The results showed
that the proposed TS-LSTME model has spatiotemporal prediction capability, which
can be applied for early warning and management of regional pollutants. We also apply
the proposed model to predict O3 concentration, and the results indicated that the model
is suitable for other air pollutants prediction.
of
predicted the average PM2.5 concentration of each station in the next 24 hours. The
hourly scale PM2.5 monitoring data came from the national city air quality real-time
publishing platform (http://106.37.208.233:20035/), and the day scale meteorological
ro
data came from the China Meteorological Data Service Center (CMDC:
http://data.cma.cn/en). The data of missing time for each kind were compensated by the
collected data of adjacent time. Figure 1 shows the distribution of air quality monitoring
-p
stations (red circles) and meteorological monitoring stations (blue triangles) in the Jing-
Jin-Ji area. The statistics of air quality monitoring stations in 13 cities are shown in
Table 1. The measured PM2.5 data at the historical moment, meteorological data, and
re
temporal data were entered into the model. Meteorological characteristics included
precipitation (PRE), air pressure (PRS), relative humidity (RHU), sunshine (SSD),
temperature (TEM), and wind direction and speed (WIN). Table 2 lists the statistical
lP
Shijiazhuang 1029A-1035A 7
Tangshan 1036A-1041A 6
Qinhuangdao 1042A-1045A 4
Handan 1047A-1050A 4
na
Baoding 1051A-1056A 6
Zhangjiakou 1057A,1059A-1061A 4
Chengde 1062A-1066A 5
ur
Langfang 1067A-1069A 3
Cangzhou 1071A-1073A 3
Hengshui 1074A-1076A 3
Jo
Xingtai 1077A-1080A 4
Where X and Y represent two variables, Cov X Y represents the covariance of two variables,
of
σ x σ y respectively represent the standard deviation of two variables X, Y.
ro
To evaluate the degree of the temporal correlation, Pearson correlation coefficient
formula was used to calculate the temporal correlations. The variables X and Y are the
-p
values of PM2.5 concentration before and after the time interval. Figure 2 shows the
changes in the Pearson correlation coefficients of the 65 air quality monitoring stations
over our Jing-Jin-Ji area for 0-50 hours intervals. It can be seen from Figure 2 that the
re
Pearson correlation coefficients decrease with the increase of the time interval. When
the time interval is less than 15h, the Pearson correlation coefficients of the front and
back time intervals are greater than 0.4. The smaller the time interval, the higher the
lP
degree of correlation.
na
ur
Jo
of
ro
-p
Figure 2. Variations among the time autocorrelation coefficients with respect to different time
lags for 65 selected stations (s1-s65).
re
Besides, there exist strong spatial correlations of PM2.5 monitoring stations. The
Pearson correlation coefficient formula was also used to calculate the correlation
coefficients among monitoring stations in each city. Table 3 shows the maximum and
lP
of
input gate 𝐼𝑡 , output gate 𝑂𝑡 , and candidate state 𝑐𝑡 are calculated through ℎ𝑡−1 and
𝑋𝑡 . Then combine the forget gate 𝐹𝑡 , the input gate 𝐼𝑡 , and the candidate state 𝑐𝑡 to
update the internal state 𝐶𝑡 at the current moment. Finally combine the internal state
ro
𝐶𝑡 at the current moment and the output gate 𝑂𝑡 to obtain the external state ℎ𝑡 at the
current moment. The forward propagation formulas of the LSTM NN are as follows:
Ft W f X t U f ht 1 b f
-p (2)
It Wi X t Ui ht 1 bi (3)
re
Ot Wo X t U o ht 1 bo (4)
lP
Ct Ft Ct 1 It ct (6)
na
ht Ot tanh Ct (7)
Where W f , Wi , Wo and Wc are the weight matrices for input vector X t at current
ur
moment. U f , U i , U o and U c are the weight matrices for external state ht 1 at last
moment. b f , bi , bo and bc are the bias vectors. and tanh are activation functions
Jo
ro
2.3.2 TS-LSTME model
Integrating hourly-scale PM2.5 concentration data and day-scale meteorological
data, and taking the spatiotemporal correlations of PM2.5 monitoring stations into
-p
account, our study regarded a city as a basic unit to establish a TS-LSTME model. The
model can perform collaborative training on the data of multiple base stations in the
re
unit city to predict the average PM2.5 concentration of multiple monitoring stations in
the next 24 hours. Figure 4 shows the framework of the TS-LSTME model. The feed-
forward process of the TS-LSTME model is as follows: (1) Considering the significant
lP
spatiotemporal correlations of PM2.5 monitoring stations in each city, the historical data
of PM2.5 concentration at monitoring stations in each city are used as the inputs of the
model; (2) The time-related characteristics of PM2.5 concentration data of monitoring
stations in each city are extracted through the bidirectional LSTM layer and the fully
na
connected layer, and the results will be the PM2.5 concentration for 0-r hours (output1)
in the future. The previous step outputs are used as the inputs for the next prediction to
predict the PM2.5 concentration for r-2 * r hours (output2) in the future. Through the
combination of a bidirectional LSTM layer and a fully connected layer, the above
ur
process performs temporal sliding prediction to form the temporal sliding block until
the PM2.5 concentration in the next 24 hours is predicted, where the value of r is
determined by the optimal time lag selected. Then, the predicted results (output1,
Jo
output2, ..., outputn) are concatenated, and the training results are passed through the
bidirectional LSTM layer, which is used as the inputs of the next process. This process
is forming the TS-LSTME unit; (3) The outputs of the previous process and auxiliary
data are integrated into the fully connected layer, which could obtain the results of the
PM2.5 concentration of each monitoring station in the next 24 hours. The auxiliary data
are composed of normalized meteorological data (precipitation, air pressure, humidity,
sunshine, temperature, wind direction, and speed), and temporal data (one-hot codes
formed by seasons and months).
In the TS-LSTME model, the rectified linear unit (ReLU) function was used as the
activation function, and dropout and regularization were used respectively in the LSTM
layer and the fully connected layer to improve the generalization ability of the model
and accelerate the speed of network training. To evaluate the effectiveness of the
proposed model, root mean square error (RMSE), mean absolute error (MAE),
coefficient of determination (R2) and normalized root mean square error (NRMSE)
were used as metric indices to evaluate model performance. These indicators are
defined as follows:
N
1
RMSE P O
2
i i (8)
N i=1
1 N
MAE = Pi Oi
N i=1
(9)
of
P O
N 2
1
i i
R2 i=1
(10)
O O
N 2
i=1 i i
ro
RMSE
NRMSE = (11)
Omax Omin
-p
where Oi is the observed air pollutant concentration, Oi is the average observed air pollutant
re
concentration, Omax is the observed maximum value of air pollutant concentration, Omin is the
observed minimum value of air pollutant concentration, Pi is the predicted air pollutant
lP
Figure 4. Network framework of the TS-LSTME model for PM2.5 concentration predictions.
The model inputs consist of historical PM2.5 concentration data of all monitoring stations in the unit
city in the dark blue dotted box and auxiliary data (meteorological data and temporal data) in the
yellow dotted box. r represents the time lag, T represents the present time, S represents the
na
dimensional data: the number of input samples, the time lag for future prediction
selection, and the number of monitoring stations in the city. The time lag is also used
as the time lag interval of each output in the TS-LSTME unit to determine the value of
Jo
r. As shown in Figure 5, according to the time lag r, the next 24 hours are grouped (T
to T + r, T + r + 1 to T + 2r + 1, etc.) to form the TS-LSTME unit, and it is worth
noting that the average concentration is predicted in the last stage (T + 24-r to T + 24).
Based on the TS-LSTME unit, the model can predict the average PM2.5 concentration
of multiple monitoring stations in the next 24 hours. According to the previous temporal
correlations analysis, the time lag will be selected from {6, 8, and 12}. Taking time lag
selection 6 as an example, the TS-LSTME unit is formed by four temporal sliding
prediction results (output1, output2, output3, output4), which are respectively the
predicted PM2.5 concentration in the next 1-6 hours, 7-12 hours, 13-18 hours, and the
predicted average PM2.5 concentration in the next 19-24 hours. Experiments with
different time lag values were made to choose the best time lag, which is the recursive
time interval of the TS-LSTME unit. The experimental results of all cities in our study
area have shown that with 12 as the time-lag input, the prediction ability of the model
was better. The experimental results of Beijing in Table 4 show that when the time lag
is selected as 12, that is, r = 12, the results showed better performance with RMSE
20.00 μg/m3, MAE 13.54 μg/m3 and R2 0.72, compared with the other values of r.
of
ro
-p
re
Figure 5. Illustration of TS-LSTME Unit. r represents the time lag, T represents the present time,
lP
S represents the monitoring station, and n Indicates the number of monitoring stations.
3. Results
`3.1 Performance of the model TS-LSTME
The study obtained the data including a total of 52,584-hour PM2.5 concentration
data and 2191-day meteorological data (January 1, 2014 to December 31, 2019) in Jing-
Jin-Ji monitoring stations, 60% of which were chosen as the training set, 20% as the
validation set and 20% as the test set. After determining the best network architecture
for the current predicting tasks, models were established in units of cities, and the
training set was used to train the current TS-LSTME models. The accuracy indicators
of the model prediction results of 13 cities in the study area are shown in Table 5 and
Figure 6. Table 5 shows that the R2 values in 13 cities were above 0.70, and the R2 value
of Hengshui city was above 0.80. Moreover, the NRMSE values are between 5.96%-
9.42%. Figure 6 shows the average values of the observed PM2.5 concentration in each
city, the MAE and RMSE values of the corresponding models. The abscissa was sorted
from largest to smallest according to the average values of urban PM2.5 concentration.
We can see that the overall trends of MAE and RMSE were also gradually decreasing,
and there existed no apparent fluctuation. In cities with high concentration of air
pollution (such as Shijiazhuang, Baoding, and Tianjin city) and cities with low
concentration (such as Chengde, Zhangjiakou city), models performed well overall,
indicating that the TS-LSTME models have good capabilities of PM2.5 concentration
predictions.
Table 5. Effects of the TS-LSTME models in different cities, using three indicators: the root mean
square error (RMSE), the mean absolute error (MAE), coefficient of determination (R2) and the
normalized root mean square error (NRMSE).
of
City RMSE (μg/m3) MAE (μg/m3) R2 NRMSE (%)
ro
Tianjin 24.37 17.69 0.72 8.04
Shijiazhuang 24.94 16.91 0.78 7.16
Tangshan 23.04 16.13 0.71 8.36
Qinhuangdao
Handan
18.79
25.33
13.19
18.08
-p 0.74
0.74
6.40
9.12
re
Baoding 28.44 20.26 0.72 5.96
Zhangjiakou 12.61 9.34 0.72 9.42
Chengde 12.04 8.77 0.73 9.02
lP
of
concentration, but the forecast results appeared "high values underestimation, low
values overestimation" phenomenon, and the intervals of the prediction results were
ro
smaller than the intervals of the observation values, which also reflects the complexity
of PM2.5 concentration predictions.
-p
re
lP
na
ur
Figure 7. Distribution of daily averaged observed PM2.5 concentration of all stations and the
corresponding predictions by TS-LSTME models for Jing-Jin-Ji area from January 1, 2019 to
December 31, 2019.
Jo
Figure 8. Correlation between the predicted and observed average values by TS-LSTME models
of
for Jing-Jin-Ji area from January 1, 2019 to December 31, 2019: (a) all test samples (b) daily
samples. The red dashed line is y = x reference line.
ro
To further explore the spatiotemporal prediction capabilities of the TS-LSTME
models and the spatial correlations of PM2.5 monitoring stations, we compared
-p
predicted spatial interpolation results of PM2.5 concentration by TS-LSTME models
with observed results at monitoring stations, and selected representative comparison
results on four days (January 1st, April 1st, July 1st, and October 1st) according to the
re
seasonal characteristics of PM2.5 concentration. From January to March and October to
December each year, the PM2.5 concentration is generally high, while from April to
September, PM2.5 concentration is generally low. Figure 9 shows the spatiotemporal
lP
distributions of the observed data and predicted data on January 1st, April 1st, July 1st,
October 1st, 2019. From Figure 9, it can be found that the heavily polluted areas in the
Jing-Jin-Ji area are mainly concentrated in cities in the southern and eastern regions.
By comparison, the predicted results on January 1st in the southern area were lower
na
than the observed values, but the predicted spatial distributions on January 1st, April
1st, July 1st, and October 1st were generally consistent with the observed spatial
distributions of PM2.5 concentration. The predicted spatial distributions were accurate,
ur
which indicated that the TS-LSTME models have strong capabilities for extracting
spatiotemporal features to realize high accuracy predictions.
Jo
of
Figure 9. Spatial interpolation results of observed values and predicted values by TS-LSTME
models on Jan 1st, 2019 (a, e), on April 1st, 2019 (b, f), on July 1st, 2019 (c, g), and on Oct 1st,
ro
2019 (d, h).
in 13 cities. The MAE values predicted by the TS-LSTME, LSTME, LSTM, MLR, and
SVR models at each monitoring station were shown in Figure 10. It can be seen from
Figure 10 that the MAE values of each monitoring station of the TS-LSTME, LSTME,
ur
LSTM models are less than the MAE values of the MLR, SVR models, indicating that
the models based on deep learning which take into account the spatiotemporal
correlations of PM2.5 monitoring stations and possess the complexity of the networks,
Jo
are more suitable for complex PM2.5 concentration prediction tasks and can obtain
prediction results of multiple monitoring stations at the same time. As can be seen from
Table 6, the R2 values of the LSTME models in each city are between 0.51-0.66, the R2
values of the LSTM models are between 0.45-0.63, and R2 values of the TS-LSTME
models are above 0.70, of which evaluation indicators are better than other models. It
is worth noting that has the R2 value of the TS-LSTME model in Hengshui city is 0.81,
and the MAE value at 1076A monitoring station is 5.98 μg/m3, which is better than the
other two models. The results show that compared with the directly predicted LSTM
models and LSTME models, the TS-LSTME models use a highly time-correlated delay
for temporal sliding prediction, which can achieve higher accuracy and stability in long-
term prediction, and add auxiliary data (meteorological data and temporal data) can
improve the prediction performances of the models.
Table 6. Effects of the different models (TS-LSTME, LSTME, LSTM), using three indicators: the
root mean square error (RMSE), the mean absolute error (MAE), and R2.
City TS-LSTME LSTME LSTM
RMSE MAE R2 RMSE MAE R2 RMSE MAE R2
Beijing 20.00 13.54 0.72 24.76 17.54 0.52 24.84 17.76 0.52
Tianjin 24.37 17.69 0.72 28.51 20.30 0.52 30.45 23.24 0.45
Shijiazhuang 24.94 16.91 0.78 29.76 20.79 0.66 30.89 21.80 0.63
Tangshan 23.04 16.13 0.71 25.87 18.02 0.52 27.15 19.89 0.47
Qinhuangdao 18.79 13.19 0.74 22.14 15.01 0.51 26.21 17.13 0.46
of
Handan 25.33 18.08 0.74 31.14 22.48 0.58 36.65 25.61 0.51
Baoding 28.44 20.26 0.72 29.34 22.12 0.66 30.67 22.49 0.62
Zhangjiakou 12.61 9.34 0.72 13.65 10.44 0.57 13.93 11.68 0.52
ro
Chengde 12.04 8.77 0.73 14.20 10.00 0.51 14.85 10.10 0.47
Langfang 20.74 14.20 0.72 23.54 16.29 0.59 24.46 17.12 0.56
Cangzhou 19.91 14.27 0.74 22.83 16.75 0.62 23.60 16.86 0.59
Hengshui
Xingtai
17.94
25.31
12.35
17.84
0.81
0.76
32.71
27.73
-p
27.09
19.38
0.41
0.68
33.13
29.68
27.21
21.74
0.39
0.65
re
lP
na
ur
Jo
Figure 10. Comparison of the results (MAE) of different models (TS-LSTME, LSTME, LSTM,
MLR, SVR) at each station in study area.
of
prediction tasks.
Table 7. Structure and prediction accuracy of multiscale air pollutant concentration predictions.
ro
Task Time lag Structure of TS-LSTME Unit RMSE MAE
(value of r) (μg/m3) (μg/m 3)
24h 12 1-13h→14-24h average 20.00 13.54
48h
72h
24
36
-p
1-24h→25-48h average
1-36h→37h-72h average
24.96
28.75
17.86
20.63
re
4. Discussion
4.1 Application of predicting O3 concentration
lP
In order to further reflect the performance of the present model for other air quality
predictions, we applied the proposed model to predict O3 concentration over Jing-Jin-
Ji area. According to the correlation analysis (Table 8), it can be known that O3 is less
relevant to other pollutants, and O3 is a critical atmospheric pollutant in summer, which
na
has become a limiting factor affecting the number of days to achieve the standard. O3
also has significant spatial and temporal characteristics, seasonal characteristics, and
will be affected by meteorological conditions, which we could apply the proposed
ur
model to predict. Data from January 1, 2014 to December 31, 2016, over Jing-jin-ji area
were input into the model to train and adjust the model’s structure to achieve the best
results. Figure 11 shows the correlation between the predicted and observed daily
Jo
average O3 values by TS-LSTME models on train set (2014-2015) and test set (2016),
respectively. The proposed model obtained high R2 values (0.96 and 0.86), indicating
strong consistency between predicted and observed values. The results show that the
proposed model is suitable for predicting multiple air pollutants.
Table 8. Pearson's coefficients between multiple air pollutants
R PM2.5 SO2 NO2 CO O3 PM10
PM2.5 1.00 0.66 0.65 0.81 -0.25 0.76
SO2 - 1.00 0.56 0.80 -0.31 0.54
NO2 - - 1.00 0.70 -0.54 0.54
CO - - - 1.00 -0.45 0.62
O3 - - - - 1.00 -0.19
PM10 - - - - - 1.00
of
ro
Figure 11. Correlation between the predicted and observed daily average O3 values by TS-LSTME
-p
models on different datasets: (a) on train dataset (2014-2015), (b) on test dataset (2016). The solid
line and dashed line are the regression line and y = x reference line, respectively.
re
4.2 Limitations and future work
In methodology, this study constructed a deep learning-based, nonlinear and
adaptive PM2.5 spatiotemporal temporal sliding prediction model by the analysis of the
lP
It is suitable for predicting multiple air pollutants such as O3, PM10, etc. Moreover, the
proposed model structure can be adjusted to predict other target variables by choosing
suitable influencing factors and time lag appropriately according to the actual situation.
ur
However, there exist some limitations of the proposed model. First, the intervals of the
prediction results are often smaller than the ranges of the actual values. There exist
cases of underestimation and overestimation, which also reflects the complexity of
Jo
of
LSTM, which can maintain the high temporal correlations. The results of the
performance evaluation and comparison with other models demonstrated that the
ro
proposed model based on the method of temporal sliding prediction could be more
effective in extracting temporal features from spatiotemporal correlations of PM2.5
monitoring stations for predicting PM2.5 concentration in long-term series. The
-p
proposed model is suitable for long-term forecasting tasks and has strong practicability.
The proposed model can be applied in other regions for multiple air pollutant
predictions such as O3, PM10, etc. By spatializing predicted results of the air quality
re
monitoring stations by the TS-LSTME model, the spatial distribution characteristics of
air pollutants in the region can be evaluated in real-time.
Based on the forecast results, the relevant departments can predict where will appear
lP
more severe pollutant pollution and adjust some measures in time to effectively control
pollution emissions according to the specific conditions of different cities, such as warn
the public to go out as little as possible to reduce exhaust emissions caused by cars
driving, supervise and control the construction site to reduce the dust caused, require
na
some industrial enterprises to reduce production to ensure the stable discharge and meet
the standards, strengthen the comprehensive management of catering industry fume and
so on. The conceptual and methodological advancement of this study could assist in
ur
joint prevention and control of regional pollutants which could promote sustainable
urban development.
Jo
Conflict of Interest
We would like to submit the revised manuscript entitled “Modeling air quality
prediction using a deep learning approach: method optimization and evaluation”, which
we wish to be considered for publication in “Sustainable Cities and Society”. No
conflict of interest exits in the submission of this manuscript, and manuscript is
approved by all authors for publication. I would like to declare on behalf of my co-
authors that the work described was original research that has not been published
previously, and not under consideration for publication elsewhere, in whole or in part.
All the authors listed have approved the manuscript that is enclosed.
References:
Antanasijević, D.Z., Pocajt, V.V., Povrenović, D.S., Ristić, M.Đ., Perić-Grujić, A.A., 2013. PM10
emission forecasting using artificial neural networks and genetic algorithm input variable
optimization. Sci. Total Environ. 443, 511–519.
https://doi.org/10.1016/j.scitotenv.2012.10.110
Ausati, S., Amanollahi, J., 2016. Assessing the accuracy of ANFIS, EEMD-GRNN, PCR, and MLR
models in predicting PM2.5. Atmos. Environ. 142, 465–474.
of
https://doi.org/10.1016/j.atmosenv.2016.08.007
Bai, Y., Li, Y., Wang, X., Xie, J., Li, C., 2016. Air pollutants concentrations forecasting using back
propagation neural network based on wavelet decomposition with meteorological
ro
conditions. Atmospheric Pollut. Res. 7, 557–566. https://doi.org/10.1016/j.apr.2016.01.004
Barzeghar, V., Sarbakhsh, P., Hassanvand, M.S., Faridi, S., Gholampour, A., 2020. Long-term trend
of ambient air PM10, PM2.5, and O3 and their health effects in Tabriz city, Iran, during 2006–
-p
2017. Sustain. Cities Soc. 54, 101988. https://doi.org/10.1016/j.scs.2019.101988
Box, E.P., Jen, M., n.d. TIME SERIES ANALYSIS: FORECASTING AND CONTROL. (Revised
Edition) by 1.
re
Engel-Cox, J., Kim Oanh, N.T., van Donkelaar, A., Martin, R.V., Zell, E., 2013. Toward the next
generation of air quality monitoring: Particulate Matter. Atmos. Environ. 80, 584–590.
lP
https://doi.org/10.1016/j.atmosenv.2013.08.016
Feng, Y., Zhang, W., Sun, D., Zhang, L., 2011. Ozone concentration forecast method based on
genetic algorithm optimized back propagation neural networks and support vector machine
data classification. Atmos. Environ. 45, 1979–1985.
na
https://doi.org/10.1016/j.atmosenv.2011.01.022
Franceschi, F., Cobo, M., Figueredo, M., 2018. Discovering relationships and forecasting PM10 and
PM2.5 concentrations in Bogotá, Colombia, using Artificial Neural Networks, Principal
Component Analysis, and k-means clustering. Atmospheric Pollut. Res. 9, 912–922.
ur
https://doi.org/10.1016/j.apr.2018.02.006
Gariazzo, C., Carlino, G., Silibello, C., Renzi, M., Finardi, S., Pepe, N., Radice, P., Forastiere, F.,
Michelozzi, P., Viegi, G., Stafoggia, M., 2020. A multi-city air pollution population
Jo
of
resolution: Part I – examination of analysis nudging and all PBL schemes associated with
LSMs in meteorological model. Atmos. Environ. 44, 523–532.
ro
https://doi.org/10.1016/j.atmosenv.2009.10.045
Kolehmainen, M., Martikainen, H., Ruuskanen, J., 2001. Neural networks and periodic components
used in air quality forecasting. Atmos. Environ. 35, 815–825.
https://doi.org/10.1016/S1352-2310(00)00385-X -p
Krishan, M., Jha, S., Das, J., Singh, A., Goyal, M.K., Sekar, C., 2019. Air quality modelling using
long short-term memory (LSTM) over NCT-Delhi, India. Air Qual. Atmosphere Health 12,
re
899–908. https://doi.org/10.1007/s11869-019-00696-7
Kumar, A., Ambade, B., Sankar, T.K., Sethi, S.S., Kurwadkar, S., 2020. Source identification and
health risk assessment of atmospheric PM2.5-bound polycyclic aromatic hydrocarbons in
lP
fractionated airborne particle-bound metals using MLR, BP-ANN and SVM analyses.
Chemosphere 180, 513–522. https://doi.org/10.1016/j.chemosphere.2017.04.015
Li, C., Hsu, N.C., Tsay, S.-C., 2011. A study on the potential applications of satellite data in air
Jo
of
Pak, U., Ma, J., Ryu, U., Ryom, K., Juhyok, U., Pak, K., Pak, C., 2020. Deep learning-based PM2.5
prediction considering the spatiotemporal correlations: A case study of Beijing, China. Sci.
ro
Total Environ. 699, 133561. https://doi.org/10.1016/j.scitotenv.2019.07.367
Pan, L., Sun, B., Wang, W., 2011. City Air Quality Forecasting and Impact Factors Analysis Based
on Grey Model. Procedia Eng. 12, 74–79. https://doi.org/10.1016/j.proeng.2011.05.013
-p
Perez, P., Reyes, J., 2006. An integrated neural network model for PM10 forecasting. Atmos. Environ.
40, 2845–2851. https://doi.org/10.1016/j.atmosenv.2006.01.010
Qi, Y., Li, Q., Karimian, H., Liu, D., 2019. A hybrid model for spatiotemporal forecasting of PM2.5
re
based on graph convolutional neural network and long short-term memory. Sci. Total
Environ. 664, 1–10. https://doi.org/10.1016/j.scitotenv.2019.01.333
Sak, H., Yang, G., Li, B., Li, W., 2016. Modeling Dependence Dynamics of Air Pollution: Pollution
lP
Feldmann, H., Renner, E., Wolke, R., 2008. A model inter-comparison study focussing on
episodes with elevated PM10 concentrations. Atmos. Environ. 42, 4567–4588.
https://doi.org/10.1016/j.atmosenv.2008.01.068
ur
Vautard, R., Builtjes, P.H.J., Thunis, P., Cuvelier, C., Bedogni, M., Bessagnet, B., Honoré, C.,
Moussiopoulos, N., Pirovano, G., Schaap, M., 2007. Evaluation and intercomparison of
Ozone and PM10 simulations by several chemistry transport models over four European
Jo
of
https://doi.org/10.1016/j.atmosenv.2016.10.016
Yang, X., Wu, Q., Zhao, R., Cheng, H., He, H., Ma, Q., Wang, L., Luo, H., 2019. New method for
ro
evaluating winter air quality: PM2.5 assessment using Community Multi-Scale Air Quality
Modeling (CMAQ) in Xi’an. Atmos. Environ. 211, 18–28.
https://doi.org/10.1016/j.atmosenv.2019.04.019
-p
Zhang, B., Jiao, L., Xu, G., Zhao, S., Tang, X., Zhou, Y., Gong, C., 2018. Influences of wind and
precipitation on different-sized particulate matter concentrations (PM2.5, PM10, PM2.5–10).
Meteorol. Atmospheric Phys. 130, 383–392. https://doi.org/10.1007/s00703-017-0526-9
re
Zhao, R., Zhan, L., Yao, M., Yang, L., 2020. A geographically weighted regression model
augmented by Geodetector analysis and principal component analysis for the spatial
distribution of PM2.5. Sustain. Cities Soc. 56, 102106.
lP
https://doi.org/10.1016/j.scs.2020.102106
Zhou, Y., Chang, F.-J., Chang, L.-C., Kao, I.-F., Wang, Y.-S., 2019. Explore a deep learning multi-
output neural network for regional multi-step-ahead air quality forecasts. J. Clean. Prod.
209, 134–145. https://doi.org/10.1016/j.jclepro.2018.10.243
na
Zhu, S., Lian, X., Wei, L., Che, J., Shen, X., Yang, L., Qiu, X., Liu, X., Gao, W., Ren, X., Li, J.,
2018. PM2.5 forecasting using SVR with PSOGSA algorithm based on CEEMD, GRNN
and GCA considering meteorological factors. Atmos. Environ. 183, 20–32.
ur
https://doi.org/10.1016/j.atmosenv.2018.04.004
Jo