CENTRAL ASIAN JOURNAL OF THEORETICAL

AND APPLIED SCIENCES

Volume: 04 Issue: 08 | Aug 2023 ISSN: 2660-5317
https://cajotas.centralasianstudies.org

Design and Analysis of Air Pollution Concentration Prediction Models

Using Transfer Learning and Recurrent Neural Networks

Mankala Satish
Assistant Professor
Dept of Computer Science Engineering (AI&ML)
CMR Institute of Technology, Kandlakoya, Hyderabad, India

Peddi Niranjan Reddy

Assistant Professor
Dept of Computer Science Engineering (AI&ML)
CMR Institute of Technology, Kandlakoya, Hyderabad, India
.

Received 24th Jun 2023, Accepted 13th Jul 2023, Online 31nd Aug 2023

Abstract: Since air pollution (AP) poses a serious risk to human health, many people have started
paying greater attention to it in recent years. Precise Air Pollution prediction helps individuals schedule
their outdoor activities and contributes to human health protection. In this study, recurrent neural
networks (RNNs) with long short-term memory (LSTM) were used to predict Macau's future APS
concentration. Data on the concentration of APS as well as environmental data have also been used.
Additionally, some air quality monitoring stations (AQMSs) in Macau have fewer overall observed data
while simultaneously collecting less observed data for specific APS kinds. In order to help AQMSs with
less observed data, transfer learning, and pre-trained neural networks have been utilized. The purpose of
this study is to show how a collection of neural network algorithms has been utilized for these two
pollutant elements. The approach is given considerable thought in this paper, and datasets regarding air
and water pollution as well as expected parameters were additionally collected for future development
efficiency.
Keywords: Transfer learning, Recurrent Neural Network

_____________________________________________________________________________________________________

I. Introduction

With the increasing development of global industrialization, air pollution has become more difficult.
Air pollution could be an important factor in infectious diseases and a decrease in life expectancy, based
to studies [1]. Countries across the world have implemented an array of methods to address this global

© 2023, CAJOTAS, Central Asian Studies, All Rights Reserved 135

Copyright (c) 2023 Author (s). This is an open-access article distributed under the terms of Creative Commons
Attribution License (CC BY).To view a copy of this license, visit https://creativecommons.org/licenses/by/4.0/
CENTRAL ASIAN JOURNAL OF THEORETICAL AND APPLIED SCIENCES
Volume: 04 Issue: 07 | Jul 2023, ISSN: 2660-5317
problem, and people have a propensity to change their routines in order to deal with the decreasing quality
of the air. Predicting air quality would be an essential and successful means for helping people in their
struggle towards such destruction. Long-term exposure to air pollution carried by vehicular traffic can
decrease life expectancy. In addition, people who are exposed to vehicle-related AP over an extended
period may live shorter lives. In numerous northern Chinese cities, Chen et al. [2] the association between
APS and lung cancer patients as well as the connection between NO2, SO2, and PM10 concentrations and
lung cancer mortality. The statistical data that they collected shows a positive correlation among the
prevalence and death of lung cancer and the level of pollutants in the air in a particular region. The air
will travel vertically upward if the atmosphere is unstable, on the contrary together, which will help in the
APS's potential to spread to the sky. The amount of APS can be viewed as time series data since LSTM
RNNs are effective at predicting time series data. Therefore, LSTM RNN has been used in this paper to
be able predict the initially indicated concentration of APS despite an absence of knowledge about the
atmospheric dispersion modeling of APS. Additionally, transfer learning has been proposed to be
employed in this study to help forecast the AP level in order to achieve decent prediction results
considering the lack of observed data.

Figure 1: India's sources of pollutants in the air pollutants

Air pollution predictions were achieved through bidirectional LSTM techniques. Recent research on
air quality prediction has concentrated on enhancing precision while maintaining the prediction window
within 12 hours; some studies have even restricted the window to 1 hour in order to achieve the highest
accuracy. Given that combating air pollution takes more time, short-term prediction has minimal practical
value despite its exceptional accuracy.

II. Related Work

In contrast with standard methods of machine learning, transfer learning approaches use knowledge
accumulated from data in additional domains to facilitate predictive modeling consisting of multiple data

© 2023, CAJOTAS, Central Asian Studies, All Rights Reserved 136

Copyright (c) 2023 Author (s). This is an open-access article distributed under the terms of Creative Commons
Attribution License (CC BY).To view a copy of this license, visit https://creativecommons.org/licenses/by/4.0/
CENTRAL ASIAN JOURNAL OF THEORETICAL AND APPLIED SCIENCES
Volume: 04 Issue: 07 | Jul 2023, ISSN: 2660-5317
patterns in the current domain. Transfer learning approaches seek to establish a framework for making use
of previously-acquired knowledge to solve new but similar problems much more quickly and effectively.
In order to continue through the timeline in transfer learning, Sinno Jian Pan and Qiang Yang published a
survey on the subject in 2009 and thoroughly examined inductive, transductive, and unsupervised transfer
learning. By applying a co-training framework, Yu Zheng et al. [3] proposed a semi-supervised learning
approach which consists of two separate classifiers: one is a spatial classifier constructed from an artificial
neural network, and the other is a temporal classifier developed on a linear-chain Conditional Random
Field (CRF). Employed auto-encoder as a trained technique to improve the performance of deep recurrent
neural network with the goal to forecast PM2.5 in Japan. The Levenberg-Marquardt approach was used to
train artificial neural networks by Asha B. Chelani et al.[4]. to determine the SO2 concentration in three
cities in Delhi.

1. Theory-Based Methods
Two theory-based air quality models which have gained popularity in the industry are the
community multiscale air quality (CMAQ) modeling system and the comprehensive air quality model
with extensions (CAMx). Both were developed using the concept of "one atmosphere," which
includes conversion and interactions among multiple kinds of air contaminants, and replicated scales
of multipollutants. The majority of airborne allergens continue to be the focus of CMAQ, which also
analyzes the general level of air quality over a number of locations. Though being generally
applicable, it has an array of intrinsic problems, including errors caused by manually defined
parameters and mass conservation being adversely affected by varying meteorological fields[5]. For
instance, the particular substance for comprehensive air quality model with extensions (PMCAMx)
predicted significantly higher levels of O3 and PM than CMAQ. Considering the fact that theory-
based models will perform well given accurate data, their effectiveness is still limited by the models'
extensive estimation and inherent little errors. In addition, it makes it tricky for them to gather a lot of
data effectively.
III. Methodology
1. Long short-term memory
Long short-term memory is a type of recurrent neural network architecture used in deep learning.
RNN solely employs feed forward neural networks to operate, whereas LSTM uses feedback
connections. The LSTM system can process complete data sequences as well as to single data points.
Due to its capacity to believe time series, this algorithm is frequently used in cases of contamination
of the water and air. We increase the constant error carrousel CEC provided by the self-connected,
linear unit j by including novel features in order to build an architecture which enables constant error
through special, self-connected units without the disadvantages of the naive approach[6]. In order to
prevent the memory content stored in j from being disrupted by unrelated inputs, a multiplicative
input gate unit is implemented. The addition of a multiplicative output gate unit similarly shields other
units from disturbance by stored, currently irrelevant memory contents.

© 2023, CAJOTAS, Central Asian Studies, All Rights Reserved 137

Copyright (c) 2023 Author (s). This is an open-access article distributed under the terms of Creative Commons
Attribution License (CC BY).To view a copy of this license, visit https://creativecommons.org/licenses/by/4.0/
CENTRAL ASIAN JOURNAL OF THEORETICAL AND APPLIED SCIENCES
Volume: 04 Issue: 07 | Jul 2023, ISSN: 2660-5317

Figure 2: Transfer learning technique.

The signal realizes the storage of data while remaining in a specific time period as it flows
through the input gate, output gate, and forget gate in turn. The neural network's LSTM structure
shows that the input variable immediately travels horizontally from input to output. In a result, the
error in prediction will begin to rise before increasing suddenly in direct proportion with the previous
occurrence in the prediction model[7]. The LSTM prediction accumulation error in the forecasting for
the regular demand for electricity. The data set used for training for the short-term load forecasting
approach usually lasts a day or an entire week.

2. Bi‑LSTM
A bidirectional LSTM is recommended here to deal with the accumulative error problem, which
is shown in Figure 3. The two layers of LSTM structure which make up the bidirectional LSTM
neural network are used to calculate the hidden vector from the front to the back and from the back to
the front, accordingly. These two layers regulate the output of the bidirectional LSTM neural network.
The standard feed-forward mechanism neural network is unique from the bidirectional LSTM neural
network. In a bidirectional LSTM, the internal nodes in each layer have no connection to one another.
With the goal to improve the association of single pieces of information in multiple-time series, a
directional loop is included in the connection of hidden layers, foregoing data; outcomes are learned
and stored in the memory unit. Combining the previous output to the current input produces the neural
network's current output. However, due to a lack of delay window width, there currently are going to
be gradient disappearance and gradient explosion issues when the time series' input data grows in
volume[8].

© 2023, CAJOTAS, Central Asian Studies, All Rights Reserved 138

Copyright (c) 2023 Author (s). This is an open-access article distributed under the terms of Creative Commons
Attribution License (CC BY).To view a copy of this license, visit https://creativecommons.org/licenses/by/4.0/
CENTRAL ASIAN JOURNAL OF THEORETICAL AND APPLIED SCIENCES
Volume: 04 Issue: 07 | Jul 2023, ISSN: 2660-5317

Figure 3: Basic bidirectional LSTM structure

The bidirectional LSTM neural network, which is based on the traditional LSTM model, can take
into consideration of the front and back correlation of the load data in time series and enhance the
model's performance, for the sequence classification problem.
The forward layer's input data sequence is used as training data during the training phase with one
another, and the backward layer serves as the reverse copy of the input data sequence. In order to
prevent the forgetting of the order information, the outcomes of bidirectional structure prediction are
impacted by the previous input and the subsequent input, enhancing the dependence between the
training data [9].
3. Recurrent neural networks
Recurrent neural networks resemble feed-forward neural networks including an addition of edges
that span recurrent time steps, providing the model an awareness of time. RNNs may not have cycles
among conventional edges, which is comparable to feed-forward networks. Recurrent edges,
connecting adjacent time steps, can ultimately generate cycles, including instances of length one that
connect a node to itself all through time. A neural network architecture referred to as RNNs can be
used to model sequence data. The behavior of RNNs, which are built from feed-forward networks, is
comparable to that of human brains. Simply put, recurrent neural networks are better than other
algorithms at predicting sequential data[10]. In conventional neural networks, all the inputs and
outputs are independent of one another. However, there are instances in which prior words are needed,
such as when predicting the next word of a sentence, and it consequently is important to remember the
prior words. RNN was developed as a result, which used a Hidden Layer to solve the issue. The
Hidden state, which retains particular details about a sequence, is the most essential component of an
RNN. RNNs have a memory that allows them to maintain a record of all the algorithm's data. Since it
produces the same output by performing the identical action on all inputs or hidden layers, it employs
identical parameters for each input.

© 2023, CAJOTAS, Central Asian Studies, All Rights Reserved 139

Copyright (c) 2023 Author (s). This is an open-access article distributed under the terms of Creative Commons
Attribution License (CC BY).To view a copy of this license, visit https://creativecommons.org/licenses/by/4.0/
CENTRAL ASIAN JOURNAL OF THEORETICAL AND APPLIED SCIENCES
Volume: 04 Issue: 07 | Jul 2023, ISSN: 2660-5317

Figure 4: Structure of Recurrent Neural Network

IV. Experimental Results

In order to better understand how LSTM, GRU, and SEQ2SEQ, performed, that represents the
expected outcome of the first and third tests. Other models have a similar weakness and periodicity. Table
1 shows the results of the experiment. Segmentation indicates the observed time span may extend up to 48
hours and that the overall time span is 72 hours, whereas the anticipated span is just 24 hours. The results
of the experiment were assessed in MSE, for pre-trained and randomly initialized networks utilizing
training and validation data. The training results for various networks, including neural networks for
various AQMSs that predict PM2.5 based on the PM10 pre-trained neural networks at all AQMSs and
neural networks for various AQMSs that predict PM2.5 based on randomly initialized neural networks for
various AQMSs, are reported. The neural networks that employed pre-trained techniques and utilized the
pre-trained network that is forecasting the same outcomes can be observed[11].
Table 1: Conduct a simulation on the dataset with a time range 48 node
Model MAE MAP COR
E
LSTM 1 90.25 0.30
2.
GRU 11.70
59 77.88 0.38
SEQ2SEQ 15.35 111.4 0.35
0
Predicted values the resulting set of graphs demonstrates that, in general, forecasts were correct when
APS concentrations were low. The problem is that outliers like the high concentration of air pollution
exist, and RNNs were not intended to consciously deal with outliers during training. In most instances,
using pre-trained neural networks were better than random initialized neural networks in terms of Best
MSE and the number of epochs necessary to obtain the Best MSE, as can be seen from the other set of
charts, which are trends of loss functions using training data and are shown in Figures.

© 2023, CAJOTAS, Central Asian Studies, All Rights Reserved 140

Copyright (c) 2023 Author (s). This is an open-access article distributed under the terms of Creative Commons
Attribution License (CC BY).To view a copy of this license, visit https://creativecommons.org/licenses/by/4.0/
CENTRAL ASIAN JOURNAL OF THEORETICAL AND APPLIED SCIENCES
Volume: 04 Issue: 07 | Jul 2023, ISSN: 2660-5317

Figure 5: Results from learning employing a 64-node dataset.

Figure 6: Real and anticipated values in a high-density residential area are contrasted.

V. Conclusion
We included a time component to reflect the "decay" effect of time on forecast because time and air
pollution prediction have a significant relationship. A hidden state decoder that may be used to obtain the
variation trend in historical and forecasted data was also provided. Moreover, we suggested a window
approach to stabilize the total number of concealed states. Pretrained neural networks that are initialized
with LSTM RNNs can increase prediction accuracy. Additionally, it is possible to decrease the number of
epochs needed to train LSTM RNNs to convergence. The novel approach gives RNNs better beginning
states. Predicting the values of the next day's pollutant concentration is the focus of our present research.

References
1. Asha B. Chelani et al, Mohammad M, Al-Sultan A (2020) A new method for prediction of air
pollution based on intelligent computation. Soft Computing 24(1):661–680
2. C. Zhang, L. Di, Z. Sun, L. Lin, E. G. Yu, and J. Gaigalas, “Exploring cloud-based web processing
service: a case study on the implementation of CMAQ as a service,” Environmental Modelling &
Software, vol. 113, pp. 29–41, 2019.
3. Asha B. Chelani et al, C. Gao et al., “Multimodel simulations of a springtime dust storm over
northeastern China: implications of an evaluation of four commonly used air quality models (CMAQ

© 2023, CAJOTAS, Central Asian Studies, All Rights Reserved 141

Copyright (c) 2023 Author (s). This is an open-access article distributed under the terms of Creative Commons
Attribution License (CC BY).To view a copy of this license, visit https://creativecommons.org/licenses/by/4.0/
CENTRAL ASIAN JOURNAL OF THEORETICAL AND APPLIED SCIENCES
Volume: 04 Issue: 07 | Jul 2023, ISSN: 2660-5317
v5.2.1 CAMx v6.50 CHIMERE v2017r4 and WRFChem(v3.9.1),” Geoscientific Model
Development, vol. 12, no. 11, pp. 4603–4625, 2019.
4. Y. Bai, Y. Li, B. Zeng, C. Li, and J. Zhang, “Hourly PM2.5 concentration forecast using stacked
autoencoder model with emphasis on seasonality,” Journal of Cleaner Production, vol. 224, pp. 739–
750, 2019.
5. J. Zhu, P. Wu, H. Chen, L. Zhou, and Z. Tao, “A hybrid forecasting approach to air quality time
series based on endpoint condition and combined forecasting model,” International
6. Journal of Environmental Research and Public Health, vol. 15, no. 9, p. 1941, 2018.

7. M. Ðuri´c and D. Vujovi´c, “Short-term forecasting of air pollution index in Belgrade Serbia,”
Meteorological Applications, vol. 27, p. 5, 2020.
8. A. Glowacz, “Fault diagnosis of electric impact drills using thermal imaging,” Measurement, vol.
171, Article ID 108815, 2021.
9. Y. Zhang, B. Chen, G. Pan, and Y. Zhao, “A novel hybrid model based on VMD-WTand PCA-BP-
RBF neural network for short-term wind speed forecasting,” Energy Conversion and Management,
vol. 195, pp. 180–197, 2019.
10. J. Chen, J. Lu, J. C. Avise, J. A. DaMassa, M. J. Kleeman, and A. P. Kaduwela, “Seasonal modeling
of PM2.5 in California’s san joaquin valley,” Atmospheric Environment, vol. 92, pp. 182–190, 2014.
11. D. R. Liu, S. J. Lee, Y. Huang, and C. J. Chiu, “Air pollution forecasting based on attention-based
LSTM neural network and ensemble learning,” Expert Systems, vol. 37, Article ID e12511, 2020.
12. Swathi, P. (2021). Industry Applications of Augmented Reality and Virtual Reality. Journal of
Artificial Intelligence, Machine Learning and Neural Network (JAIMLNN) ISSN: 2799-1172, 1(02),
14-18.
13. Swathi, P. (2013). Scope of Financial Management and Functions of Finance. International Journal of
Advanced in Management, Technology and Engineering Sciences. ISSN NO-2249-7455, 3, 109-116.
14. Swathi, P. (2022). Implications For Research In Artificial Intelligence. Journal of Electronics,
Computer Networking and Applied Mathematics (JECNAM) ISSN: 2799-1156, 2(02), 25-28.
15. Swathi, P. (2022). A Study On The Restrictions Of Deep Learning. Journal Of Artificial Intelligence,
Machine Learning and Neural Networks, ISSN-2799-1172, 2(2), 57-61.
16. Bhaskar, N., Ramana, S., & Kumar, G. M. (2023, January). Internet of Things for Green Smart City
Application Based on Biotechnology Techniques. In 2023 International Conference on Artificial
Intelligence and Knowledge Discovery in Concurrent Engineering (ICECONF) (pp. 1-7). IEEE.
17. Ramana, S., Bhaskar, N., Murthy, M. R., & Sharma, M. R. (2023). Machine Learning for a Payment
Security Evaluation System for Mobile Networks. In Intelligent Communication Technologies and
Virtual Mobile Networks (pp. 347-356). Singapore: Springer Nature Singapore.
18. Ramana, S., Bhaskar, N., & Murthy, M. R. (2020). Three Level Gateway protocol for secure M-
Commerce Transactions. Solid State Technology, 63(6), 11155-11174.
19. Pothuganti, K. (2021). Long Short-Term Memory (LSTM) algorithm based prediction of stock
market exchange. International Journal of Research Publication and Reviews, 2(1), 90-93.
20. Pothuganti, K. (2021). Long Short-Term Memory (LSTM) algorithm based prediction of stock
market exchange. International Journal of Research Publication and Reviews, 2(1), 90-93.

© 2023, CAJOTAS, Central Asian Studies, All Rights Reserved 142

Copyright (c) 2023 Author (s). This is an open-access article distributed under the terms of Creative Commons
Attribution License (CC BY).To view a copy of this license, visit https://creativecommons.org/licenses/by/4.0/