Delhi air quality prediction using LSTM deep learning models with a focus on COVID-19

lockdown

Animesh Tiwaria , Rishabh Guptab , Rohitash Chandrac

a Department of Civil Engineering, Indian Institute of Technology Guwahati, Guwahati, India
b Department of Geology and Geophysics, Indian Institute of Technology Kharagpur, Kharagpur, India
c School of Mathematics and Statistics, UNSW Sydney, NSW 2006, Australia

Abstract
Air pollution has a wide range of implications on agriculture, economy, road accidents, and health. In this paper, we use novel
arXiv:2102.10551v1 [cs.LG] 21 Feb 2021

deep learning methods for short-term (multi-step-ahead) air-quality prediction in selected parts of Delhi, India. Our deep learn-
ing methods comprise of long short-term memory (LSTM) network models which also include some recent versions such as
bidirectional-LSTM and encoder-decoder LSTM models. We use a multivariate time series approach that attempts to predict air
quality for 10 prediction horizons covering total of 80 hours and provide a long-term (one month ahead) forecast with uncertainties
quantified. Our results show that the multivariate bidirectional-LSTM model provides best predictions despite COVID-19 impact
on the air-quality during full and partial lockdown periods. The effect of COVID-19 on the air quality has been significant during
full lockdown; however, there was unprecedented growth of poor air quality afterwards.
Keywords: Deep learning, LSTM, air pollution, Delhi, COVID-19, prediction

1. Introduction shows that infected air is one of the prominent causes of pre-
mature deaths [14] and the average life span reduces due to in-
The global human population has risen by more than four creasing rates of air contamination [15]. It is well known that
times during the last century [1]. Research show that the major apart from industrial pollution, agricultural fires are one of the
growth in population is attributed to the metropolitan areas in significant contributors of air pollution in Delhi [16].
the less developed regions around the globe [2, 3]. The conse- Apart from strategies that focus on management of industrial
quence of these increased levels of growth in poorly developed waste[17], it is important to model and forecast the air-quality
states is low air quality. We note that 80% of global cities [3] both in short and long-term. Machine learning methods have
and 98% of cities in middle-income countries surpass the pro- been promoting for forecasting temporal sequences and their
posed levels of air quality [4, 5]. Increase in air pollution results application to air-quality forecasting has gained attention re-
in economic losses, reduced visibility, contributes to faster cli- cently [18, 19, 20]. Forecasting models can be used to develop
mate change that contribute to extreme weather conditions, mil- strategies to evaluate and alarm the general public for future
lions of premature deaths annually [6]. The major factor in air hazardous levels of air quality index. Forecasting models for
pollution is the anthropogenic fine particulate matter (PM); i.e. air pollution concentrations can be broadly classified into two
PM2.5 (particles with an aerodynamic diameter shorter than 2.5 major categories; simulation based and data driven approaches
micrometer) [7, 8, 9, 10]. Despite the concentrations of PM2.5 such as statistical or machine learning methods. Simulation-
being two to five times higher in developing countries, most of based method incorporates physical and chemical models for
the air quality considerations and estimations are analyzed for generating meteorological and background parameters to simu-
developed countries [11]. late emission, transport and chemical transformation of air pol-
Delhi is one of the most prominent in terms of growing cities lution [21, 22]]. However, they suffer from numerical model
has an estimated population of more than 19.3 million [12]. The uncertainties and due to the lack of data, the parameterization
population density and growth in the last few decades and rapid of aerosol emissions is restricted [23]]. Data driven approaches
industrial expansion led to massive air pollution to hazardous exploit statistical and machine learning techniques to detect pat-
levels and thus failed in providing people with one of the pri- terns between predictors and dependent variables in temporal
mary life amenities, quality air. The World Economic Forum sequences [24, 25, 26, 27, 28]. Machine learning methods can
recently reported India having 6 of the world’s 10 most pol- be used to identify the exposures relevant to health outcomes of
luted cities with Delhi has one of most polluted [13]. Research interest within high-dimensional data [29]. Advances in deep
learning methods give further motivations for application to do-
main of air-quality prediction.
Email addresses: [email protected] (Animesh Tiwari),
[email protected] (Rishabh Gupta), Recurrent neural networks (RNNs) are prominent deep learn-
[email protected] (Rohitash Chandra) ing models [30] suited for modelling temporal sequences, espe-
Preprint submitted to Environmental Modelling & Software February 23, 2021
cially those involving long-term dependencies [31, 32, 33, 30, 2. Related Work
34]. Long Short Term Memory networks (LSTMs) were de-
veloped [33] to address limitations in learning long-term de- 2.1. COVID-19 and air pollution
pendencies in sequences by canonical RNNs [35, 36]. Gated
Recurrent Unit (GRU) [37, 38] networks are more simpler to There have been studies relating the effect of environment
implement but provides similar performance than LSTMs. Bi- and air pollution on COVID-19 and vice versa [48, 49, 49].
directional RNNs connect two hidden layers of opposite direc- Zhu et al. [50] found that there was a significant relationship
tions to the same output where the output layer can get informa- between air pollution and COVID-19 infection in China. Ker-
tion from past and future states simultaneously [39]. This led imray et al. [48] presented an assessment on the impact of
to bidirectional-LSTMs for phoneme classification [40] which COVID-19 in large cities in Kazakhstan and found that the tem-
performed better than standard RNNs and LSTMs and has the poral reduction in pollution may not be directly attributed to the
potential to be used for forecasting air-pollution time series. lockdown due to favorable meteorological variations during the
The coronavirus disease 2019 (COVID-19) is an infectious period, however the spatial effects of the lockdown on the pollu-
disease [41, 42, 43] which became a global pandemic [44] with tion levels were clear. Moreover, other non-traffic based sources
major impact to our lifestyle. COVID-19 forced many countries such as coal based power plants substantially contributed to the
to close their borders and enforce a partial or full lock down pollution level. Dantas et al. [49] presented a study on the im-
with devastating impact on the world economy [45, 46, 47]. pact of COVID-19 partial lockdown on the air quality of the
There has been studies relating the effect of environment and air city of Rio de Janeiro, Brazil. The authors reported that The
pollution on COVID-19 and vice versa [48, 49, 49] with studies carbon-dioxide levels showed the most significant reductions
regarding China [50], Kazakhstan [48] and [49] Brazil. In some during the partial lockdown while nitrogen-oxide decreased in
cases, COVID-19 lock-downs with reduced traffic showed to re- a lower extent, due to industrial and diesel input. The air qual-
duce air pollution while in others, it did not make significant im- ity index (with PM-10 concentration) was only reduced during
pact due to meteorological conditions and factors such as indus- the first partial lockout week. Bao and Zhang [53] reviewed 44
trial pollutants [51, 49]. Although machine learning has been cities in China to find if COVID-19 lockdowns had an effect
used for forecasting air-quality, there is scope in improving the on air pollution. The authors showed that the lockdowns of 44
forecasts in using latest machine learning models, that feature cities reduced human movements by 69.85 %, and the reduc-
deep learning models such as recurrent neural networks. Given tion in the air quality index was mediated by human mobility.
this motivation, we focus on air-quality index of Delhi which Li et al. [54] presented a study on air quality changes during
lately reached hazardous levels. We note that the air-quality the COVID-19 lockdown over the Yangtze River Delta region
in Delhi has significantly improved post COVID-19 pandemic in Northern China where the authors showed that the ozone did
[52], however this was due to lock-downs by the government not show any reduction. Moreover, even during the lockdown
and the air-quality can deteriorate further when the restrictions it was evident that background and residual pollution are still
are eased. Hence it is important to develop robust forecasting high, which includes sources mostly from the industry. Dutheil
models, that are applicable even during lock-downs and eras of and Naval [55] investigated COVID-19 as a factor influencing
lock-downs during the COVID-19 pandemic. air pollution for China. The authors argued that COVID-19
In this paper, use novel deep learning methods for long-term pandemic might paradoxically have decreased the total number
air-quality prediction in selected parts of Delhi, India. Our of deaths during the period by drastically decreasing the num-
deep learning methods comprise of Long Short Term Memory ber of fatalities due to air pollution, apart from positive benefits
(LSTMs) networks, bidirectional-LSTMs and encoder-decoder in reducing preventable non communicable diseases. Wang et
LSTMs. We use a multivariate time series approach that at- al. [51] showed that in case of China, severe air pollution events
tempts to predict air quality for 10 prediction horizons covering were not avoided by reduced activities during COVID-19 out-
a total of 80 hours with 8 hours for each prediction horizon. We break due to adverse meteorological events. The authors high-
also provide a long-term (one month ahead) forecast with un- light that large emissions reduction in transportation and slight
certainty quantification given feedback of predicted values into reduction in industrial would not help avoid severe air pollution
the model. We first show visualisation of the air quality indi- especially when meteorology is unfavorable.
cators before and during coronavirus (COVID) restrictions in There have been studies if there is another correlation with
Delhi. We investigate the impact of COVID-19 on air-quality rise in COVID-19 infections to the weather change. Tosepu et
and the ability of the model to provide quality predictions be- al. [56] presented a study regarding the correlation between
fore and during COVID-19. Our models feature data that con- weather and COVID-19 pandemic in Jakarta, Indonesia tak-
siders major seasons, effect of COVID-19, and considers mul- ing into account different temperature levels and percentage of
tivariate and univariate approach for predictions. humidity, and amount of rainfall. The authors reported that
The rest of the paper is organised as follows. Section 2 only the average temperature was significantly correlated with
presents a background and literature review of related work. COVID-19 pandemic. Ma et al. [57] presented a study on the
Section 3 presents the proposed methodology and Section 4 effects of temperature variation and humidity on the death of
presents experiments and results. Section 5 provides a discus- COVID-19 in Wuhan, China. The authors collected the daily
sion and Section 6 concludes the paper with discussion of future death numbers of COVID-19 and meteorological parameters
work. and air pollutant data and used generalized additive model to
2
explore the effect of temperature, humidity and diurnal temper- to air pollution during the daily commute in a systematic re-
ature range on the daily death counts of COVID-19. The au- view highlighted studies with schoolchildren’s exposure during
thors reported that the temperature variation and humidity may commutes that are linked with adverse cognitive outcomes and
be important factors affecting the COVID-19 mortality. severe wheeze in asthmatic children[67]. Furthermore, Ambi-
ent air pollution associated with reduction in lung functionality
2.2. Machine learning methods for forecasting air pollution and other respiratory conditions among children [68].
Apart from health, air pollution has taken a toll on various
We review machine learning methods used for air quality
other sectors, which includes agriculture. Industrial air pollu-
forecasting in the past decades. Machine learning methods
tion has a drastic effect in agricultural production as shown by
have achieved tremendous success in a variety of areas for air
a study in China with lower marginal products and further al-
quality forecasting [24, 25, 26, 27, 28]. Although neural net-
ters the relationships of labour-capital and other factors [69].
works have advantages over traditional statistical methods in air
Finally, air pollution has a drastic effect on development and
quality forecasting, they have room for improvements [58] due
economy. A study on the relationship between air pollution and
to challenges that include computational expense, sub-optimal
stock returns further showed that industrial air pollution signif-
convergence, over-fitting, and noisy data. Moreover, the chal-
icantly reduces the technical efficiency of agricultural produc-
lenge is in configuration of the network topology and model
tion [70].
parameters which affects the prediction performance. Corani
[59] used models to predict hourly PM-10 concentrations on
the basis of data from the previous day with feedforward and 3. Methodology
pruned neural networks. Jiang et al. [? ] explored multiple
models that feature physical and chemical model, regression 3.1. Delhi Situation Report and Data
model, and multiple layer perceptron on the air pollutant pre-
According to the United Nations 2016 report, Delhi had an
diction task, and their results show that statistical models are
estimated population of 26 million in greater metro area which
competitive with the classical physical and chemical models.
is projected to rise to 36 million by 2030 and will remain sec-
Machine learning is an active field of research and every
ond most populated city after Tokyo [71]. The current popula-
now and then novel tools and techniques emerge for more re-
tion density is also one of the worlds largest that will continue
fined modeling of a specific problem. Some recent works use
to pose further challenges to air quality and health. According
straightforward approaches like box models, Gaussian models
to a study by Indian Ministry of Earth Sciences in 2018, it was
and linear statistical models, which are easy to implement and
shown that number of vehicles increases by four times since
allow for the rapid calculation of forecasts [60]. Fu et al. [61]
2000 which has been a major factor for air pollution in India
applied a rolling mechanism and gray model to improve tradi-
which includes PM2.5 and hazardous nitrogen-oxide [72]. In
tional neural network models. Chang et al. [62] used aggre-
the past five years, the air quality index of Delhi has been gen-
gated LSTM networks and compared the results with support
erally moderate (101–200) level between January to Septem-
vector regression, and gradient boosted tree regression. Karim-
ber. The air quality index then drastically deteriorates to very
ian et al. [23] implemented three machine learning methods to
poor (301–400), and further Severe (401–500) or Hazardous
forecast air quality given by PM-2.5 concentrations over dif-
(500+) levels during October to December due to various fac-
ferent time intervals; these include multiple additive regression
tors [73, 74]. The air pollution status in Delhi has undergone
trees, deep feedforward neural network, and a hybrid model
many changes in terms of the levels of pollutants and the control
based on LSTM networks, where the LSTM model was most
measures taken to reduce them. Sulian et al. [75] provides an
effective for forecasting and controlling air pollution. Xiao et al.
evidence-based insight into the status of air pollution in Delhi
[63] parameterised non-intrusive reduced order model based on
and its effects on health and control measures instituted.
proper orthogonal decomposition for model reduction of pol-
The meteorological conditions, such as regional and synop-
lutant transport equations to provide rapid response urban air
tic meteorology are significant in determining the air pollutant
pollution predictions and controls. Zhu et al. [64] focused on
concentrations[76, 77, 78, 79]. Lower wind speed (weak dis-
refined modeling for predicting hourly air pollutant concentra-
persion/ventilation) can result in higher concentrations of traffic
tions on the basis of historical meteorological and air pollution
pollutants [80]. However, strong wind speed might form dust
data.
storms and end up blowing the particles on the ground.[81].
Higher humidity levels are associated with higher aggregates
2.3. Effect of air-pollution of air pollutants like PM2.5, carbon monoxide (CO), nitrogen
Data-driven machine learning methods present an opportu- dioxide (NO2) and sulfur dioxide (SO2) [82]. In addition, high
nity to simultaneously assess the impact of multiple air pol- levels of humidity often indicate precipitation events which
lutants on health outcomes. There is growing evidence that result in heavy wet deposition leading to the lowering of air
early-life exposure to ambient air pollution affect neuro devel- pollutants[83]. The most important factors of attenuated visibil-
opment in children [65]. A study that showed that air pol- ity are the interactions of particle compositions and light [84?
lution is a potential risk factor for obesity with higher body- ]. Thus, low visibility can be considered as a strong indicator
mass index in adults that warrants further investigation about of high PM2.5 concentrations. Many a time, the formation of
other health effects[66]. Assessing schoolchildren’s exposure some significant air pollutants like ozone (O3) is reduced by
3
cloud cover as they absorb and scatter solar radiations [85, 86].
Therefore, these meteorological variables are importantly se-
lected to predict air quality and hence we take them into con-
sideration.
We chose four most polluted areas in Delhi and several sur-
rounding districts known as the National Capital Region (NCR)
for our study. Bawana was the most-polluted area with an air
quality index of 497, followed by DTU-Delhi Technological Parameter Range Mean and std
University (487), Anand Vihar (484) and Vivek Vihar (482) as PM10 13.00-1000.0 218.05 138.61
shown in Figure 3. There are 12 parameters that have been Benzene 0.03-40.86 4.96 4.20
considered for our study, the data is taken from the Central Pol- Toluene 10.37-238.50 43.02 9.18
lution Control Board, India [87] for 3 years in 8 hour interval. Ammonia Gas(NH3) 0.18-300.65 25.84 20.60
The dataset consists of concentration levels of different param- Nitric Acid(NO) 0.79-351.21 18.50 23.56
eters of air-pollution from four different monitoring stations in- Nitrous Oxide(NO2) 2.00-265.82 38.45 26.99
Nitrogen Oxides(NOx) 0.0-344.35 34.42 29.82
stalled by the government agency in Delhi-NCR region. Tables
Wind Speed(WS) 0.23-6.24 0.69 0.37
[1, 2, 3, 4] show the range of values for different parameters in Ozone 1.66-186.68 51.60 30.84
our dataset between 1 January,2019 and 10 December, 2020 for Sulphur Dioxide(SO2) 0.52-78.14 10.12 7.81
different monitoring stations. We find the monitoring station Carbon Monoxide(CO) 0.0-7.54 0.90
at Bawana records the highest range of values and mean for PM2.5 9.17-948.29 111.29 98.01
PM2.5 concentration as compared to other monitoring stations.
Table 3: Details about features for DTU.

Parameter Range(Min- Mean and std

Max)
PM10 14.19-939.30 238.95 158.72
Benzene 0-30.82 3.86 3.41
Toluene 0-450.98 32.06 40.87
Ammonia Gas(NH3) 0.10-136.30 44.78 23.78
Nitric Acid(NO) 0.3-490.90 80.60 78.38
Nitrous Oxide(NO2) 0.43-360.51 74.19 41.04
Nitrogen Oxides(NOx) 0.13-472.48 107.19 80.08
Wind Speed(WS) 0.3-4.43 0.86 0.68
Ozone 1.42-180.38 36.15 26.34
Sulphur Dioxide(SO2) 0.50-94.36 13.57 9.50
Carbon Monoxide(CO) 0.01-7.21 2.14 1.10
PM2.5 6.93-936.33 116.83 106.76

Table 1: Details about features for Anand Vihar.

Parameter Range Mean and std
PM10 50.00-994.00 211.11 137.77
Benzene 0.12-22.54 4.43 3.52
Toluene 3.17-340.20 44.48 42.43
Ammonia Gas(NH3) 4.40-206.48 33.80 22.55
Nitric Acid(NO) 0.38-440.74 18.19 37.06
Parameter Range Mean and std Nitrous Oxide(NO2) 0.20-130.12 33.49 19.25
PM10 6.88-941.00 227.86 146.87 Nitrogen Oxides(NOx) 4.29-428.66 32.73 35.49
Benzene 0.1-0.43 0.26 0.01 Wind Speed(WS) 0.20-15.0 1.69 1.35
Toluene 1.0-354.83 137.58 26.89 Ozone 0.31-180.5 39.63 29.00
Ammonia Gas(NH3) 0.03-127.2 36.45 15.99 Sulphur Dioxide(SO2) 2.08-108.51 22.14 11.63
Nitric Acid(NO) 0.27-180.95 10.28 18.56 Carbon Monoxide(CO) 0.0-7.84 1.56 0.97
Nitrous Oxide(NO2) 5.37-230.82 39.91 25.83 PM2.5 3.54 907.00 109.69 104.16
Nitrogen Oxides(NOx) 4.32-184.36 29.62 24.63
Wind Speed(WS) 0.3-4.11 1.18 0.55 Table 4: Details about features for Vivek Vihar.
Ozone 0.1-170.25 44.81 35.06
Sulphur Dioxide(SO2) 0.56-58.02 12.76 8.29
Carbon Monoxide(CO) 0.0-6.90 1.22 0.80
PM2.5 10.73- 994.00 123.37 109.30

Table 2: Details about features for Bawana.

4
Figure 1: Framework diagram for selected monitoring base stations and selected pollutants showing the data collection from base stations, pre-processing and
feeding of pre-processed data into the deep learning models for predicting PM2.5 values with uncertainty. We use three deep learning models that include, LSTM,
BD-LSTM and ED-LSTM for multivariate and univariate models as shown.

3.2. LSTM network models RNNs are simply two independent RNNs together in a structure
Recurrent Neural Networks (RNNs) are a class of artificial that allows the networks to have both backward and forward
neural networks which gained a lot of popularity in recent years. information about the sequence at every time step.
The Elman RNN [31, 88] is one of the earliest architectures and The combination of bidirectional RNNs with LSTM model
considered a prominent example of simple recurrent networks gives us bidirectional LSTM model (BD-LSTM), which have
trained by back-propagation through-time (BPTT) algorithm the ability to access long-range context in both input direc-
[32]. In the early days, there was much research for model- tions [95]. BD-LSTM are designed for specific input sequences
ing dynamic systems using simple RNNs and it was shown that whose starting and ending are known beforehand. They take
RNNs perform better than feed-forward networks in knowledge both the future and past states of each element of a sequence
representation tasks [89, 90, 91]. into consideration where one LSTM processes the information
Backpropagation through time (BPTT) [32] features error from start to end of the sequence and the other from end to
backpropagation which uses the idea of training RNNs using start. Their combined outputs predict the corresponding labels
gradient descent in a way similar to feedforward neural net- if available at each time step [30]. This approach differs from
works. The major difference is that the error is back-propagated unidirectional approach in a way where information from the
for a deeper network architecture that features states defined by future and past LSTM model states combined; hence, it pre-
time; however, BPTT experiences the problems of vanishing serves information from both past and future at each time step.
and exploding gradients in case of learning tasks which involve BD-LSTM have been prominent in sequence processing prob-
long-term dependencies or deep network architectures. [92]. lems such as phoneme classification [95], continuous speech
Long Short Term Memory(LSTM) network [33] was developed recognition [96], speech synthesis [97] and sentence classifica-
which was capable of overcoming the fundamental problems of tion [98].
deep learning [30]. LSTM networks addressed the issue with
much better capabilities in remembering the long-term depen- 3.4. Encoder-Decoder LSTMs
dencies using memory cells and gates for temporal sequences.
A sequence to sequence model lies behind numerous systems
The memory cells are trained in a supervised fashion using an
which we face on a daily basis [99], and such models aim to
adaptation of the BPTT algorithm that considers the respec-
map a fixed-length input with a fixed-length output, where the
tive gates [33]. More recently, Adam optimiser which features
length of the input and output may differ. In multi-step and mul-
adaptive learning rate has been prominent in training LSTM
tivariate analysis, both the input and output are variable with
models via BPTT [93].
potentially different lengths. This problem is analogous to ma-
chine translation between natural languages, where a sequence
3.3. Bi-directional LSTMs of words in the input language is translated to a sequence of
One shortcoming of conventional RNNs is that they are only words in the output language. Recently, it has been shown
able to make use of previous context in a sequence to predict how to effectively address sequence to sequence problem with
future states. Bidirectional RNNs [94] do this by processing the encoder-decoder LSTM networks (ED-LSTM).
data in both directions with two separate hidden layers, which ED-LSTM handles variable length inputs and outputs by first
are then feed forwards to the same output layer. Bidirectional encoding a given input sequence Xip = (x1 , x2 , ..xn ), one at a
5
Figure 2: Diagram showing the conversion of 2-Dimensional(2-D) data to 3-Dimensional(3-D) data using M windows of size N and creation of feature set for both
univariate and multivariate model to give multi-step ahead predictions. For 2-D data, the window is highlighted in blue while the future prediction feature (PM2.5)
is highlighted in pink.

time using a latent vector representation for input of length n framework diagram for our experimental setup. We pre-process
and the output is a sequence Y = (y1 , y2 , ..ym ) of length m. In the dataset extracted from four different monitoring stations
the encoding stage, ED-LSTM creates a sequence of hidden in Delhi (Figure 3 with map adopted from website 2 ) for our
states for input sequence and then in decoding stage it defines proposed deep learning framework as shown in Figure 1. Af-
a distribution for output states given the input sequence. The ter collecting the data, we fill the missing values of a particu-
main goal of the model is to maximize the conditional proba- lar feature by the median of its available neighboring values.
bility P(Y|Xip ) of mapping input sequence to output sequence The missing values may be due to inconsistency in recording
while training. ED-LSTM have been previously used in a lot the pollutant values at different monitoring stations. In or-
of tasks, some of which include text simplification [100], auto- der to train LSTM models, we transform the two-dimensional
matic speech recognition [101] and grapheme to phoneme con- (2D) dataset of features and temporal sequences given by time-
version [102]. interval (eight hours) into a into a three-dimensional (3D) for-
mat [103] by a sliding window technique in which a window of
size N moves over the original 2D dataset of size (r,c); where
r represents the number of rows(features) and c represents the
number of columns in dataframe. This results in M windows fi-
nal dataset with dimensions (M,N,r) which is fed into the mod-
els to give Nout multi-step ahead prediction values as shown
in Figure 2. We break the dataset into a feature set Xi =
(x1 , x2 , x3 ....xr−1 ), i = 1...N for multivariate model and Xi = (xr )
for univariate model and labels Y j = (y1 , y2 , ..., yNout ), j = 1...M
as multi-step ahead predictions. The label value consists of PM
2.5 value and other meteorological values constitutes the fea-
ture vector for multivariate model while for univariate model
past PM2.5 values constitute the feature vector .
Our framework for the respective LSTM models is as fol-
lows. We use N f = 11 features (different pollutant values other
than PM2.5 as shown in Table 1) for our multivariate model
where we consider N = 5 a window that captures the number
of time steps in the past. The window determines how much
the LSTM model will unfold in time and the number of fea-
tures determines the number of input neurons in the respective
LSTM models. We define Nout = 10 as the number of output
Figure 3: Map of Delhi highlighting the location of four monitoring base sta- neurons in the LSTM model which denotes the number of steps
tions in Delhi. (Note: The map is not to scale. )
ahead to predict PM2.5 concentration. We note that each step
ahead is denoted by 8 hour intervals and hence the respective
models would predict 80 hours ahead. Hence, we use 40 hours
3.5. Framework (5 steps) of past information to predict 80 hours (10 steps) of
The air-pollution dataset for four different monitoring sta- future trend of the time series. In the case of univariate model,
tions has been collected from Central Pollution Control Board we only use PM2.5 as a feature to predict its future trend and
(CPCB) [87], Government of India 1 . Figure 1 shows the hence, the respective LSTM models will feature 1 input neuron

1 https://cpcb.nic.in/ 2 http://gismaps.in/prod/indiabasemaps/Delhi/Delhi.html

6
and 10 output neurons. COVID-19 lockdown. We test the model on respective months
for 2020 which corresponds to the COVID-19 lockdown and
partial lockdown periods, in this case we use 50 epoch of train-
4. Experiments and Results
ing with our best LSTM model. Second, we use a univariate
In our experiments, we compare the performance of three time-series approach for multi-step ahead prediction of PM2.5
LSTM models as presented earlier for univariate and multivari- values and test the performance using our best model. In this
ate prediction models. We first analyse the impact of COVID- case, we use 1000 epochs of training with our best LSTM
19 lock-downs on the air quality and then evaluate the different model. In all our experimental setup using different LSTM
models for multi-step ahead prediction of air quality (PM2.5). models and FNN, we further split the test dataset into validation
Our experiments consider different data setup that features data dataset for tuning the model during trial experiments and test
before and after COVID-19 in order to evaluate the effect on air our model on the remaining portion for evaluating the models,
quality and model accuracy. we split the test dataset in the ratio 1:1 for for all our respective
experiments.
4.1. Experiment setup The prediction performance is measured by root mean
squared error (RMSE) as follows
The visualisations and experiments are organised as follows.
v
u
N
t
• Provide analysis of concentration of PM2.5 for different 1 X
RMS E = (yi − ŷi )2
monitoring stations and for different time intervals includ- N i=1
ing the COVID-19 lockdown;
where, yi and ŷi are the actual value and the predicted value,
• Compare the performance of the three LSTM network respectively. N refers to the total length of the data. We report
models using Adam optimizer; the RMSE for different horizons and also the mean RMSE for
• Create training data with different strategies, that includes all prediction horizons.
and exclude COVID-19 time-span and compare the per- We conduct 30 independent experimental runs with different
formance of best LSTM model; weight initialisation in the respective LSTM models and pro-
vide mean and standard deviation in all our results.
• Compare univariate and multivariate approach using the
best LSTM model; 4.2. Data visualisation
• Provide one month forecast using best LSTM model We first provide data visualisation and analyse the effect of
with uncertainty quantification with multiple experimen- COVID-19 on air pollution in Delhi while comparing with pre-
tal runs. vious years. Figure 5 presents PM2.5 values of stations in
Anand Vihar, Bawana, Delhi Technical University and Vivek
The technical details of our framework for the respective Vihar from 1 January 2018 to 10 December 2020 that is used
LSTM models as follows. We considered different number of as dataset for this study. We find some missing data in the ini-
hidden neurons and other hyper-parameters (such as learning tial months of 2018 for some stations therefore we consider the
rate) in our trial experiments. We define the topology of differ- data from 1 January 2019 to 10 December 2020 for building our
ent models in terms of input, hidden layers and output in Table respective models and strategies. Table 6 presents a summary
5. We use Adam optimizer [93] with batch size of 20 for 200 of the visualisation in Figure 5 where the seasons are quantita-
epochs and rectifier linear unit (ReLu) activation in the hidden tively highlighted. We observe higher values of PM2.5 value for
layers for the three LSTM models. We also compare our results the months from October till February on a seasonal basis af-
with FNN which uses Adam optimiser. ter which the value starts to descend starting from the month of
In the respective experiments, we compare the performance March. Although we find lower values of PM2.5 during March-
of different models for different time steps by using the train June every year, we observe a significant decrease for months
data from Jan 2019 to May 2020 (pre-COVID-19 lockdown) of interest (March-June 2020), which belongs to the COVID-19
and the test data from June 2020 to Dec 2020 (during and par- full lockdown period in India when compared to the respective
tial COVID-19 lockdown). We test our models by training with- months in 2018 and 2019as shown in Table 6 and Figure 5.
out shuffling the time-series data window and also by shuffling We find a similar decrease of PM2.5 values for all the monitor-
it. Shuffling the data here refers to randomly picking different ing stations while Bawana records the highest values for 2019
training data windows (regions) comprising of past time steps and 2020 among other stations. The partial lockdown period
and its corresponding output time steps. Hence, with shuffling, include the months after June, 2020 as highlighted in Figure 5.
the training data would feature peak and off-peak regions of Figure 4 shows Pearson’s correlation heatmaps for different
COVID-19 as different weeks can be featured in training data features in our dataset for different monitoring stations. Higher
rather than taking consecutive weeks. positive values of Pearson’s correlation between two particu-
Furthermore, we test the best model in two different ways, lar feature indicates a high positive dependence between them
first by training it on a seasonal data comprising of observa- while a higher negative value indicates a high negative depen-
tions from February to September in 2019 which excludes the dence between them and the values closer to zero indicates
7
 Input Dimension Hidden Layers Output Comments
FNN N × N f = 55 2 10 Hidden Layer sizes (h1 , h2 )= (64,32)
LSTM (N, N f ) = (5, 11) 1 10 Hidden Layer of 50 cells.
BD-LSTM (N, N f ) = (5, 11) 1 10 Forward and Backward layer of 50 Cells each.
ED-LSTM (N, N f ) = (5, 11) 4 10 Two LSTM layers of 50 cells each, Repeat Vector and a Time
distributed layer

Table 5: Topology of different deep learning models.

(a) Feature correlation heatmap for Anand Vihar. (b) Feature correlation heatmap for Bawana.

(c) Feature correlation heatmap for DTU. (d) Feature correlation heatmap for Vivek Vihar.

Figure 4: Correlation via heat-map for different features in the data collected from different monitoring stations in Delhi.

more independent feature sets. Higher positive correlation val- dependence. We also observe high positive correlation values
ues are represented as dark brown panels while higher nega- between PM2.5 and PM10 for all the monitoring stations which
tive correlation values are represented as dark blue panels in also indicates a good amount of correlation.
different correlation heatmaps. We observe higher correlation
values of NOx with NO and NO2 also which is expected be-
cause they are nitrogenous pollutants and show higher positive

8
 Figure 5: PM2.5 values of Anand Vihar, Bawana, Delhi Technical University, and Wazirpur from 1 Jan 2018 to 10 Dec 2020.

Table 6: Mean with +/- 95% confidence interval for PM2.5 concentrations for March-June for three consecutive years over different monitoring stations.

Time Interval PM2.5(Bawana) PM2.5(Anand Vihar) PM2.5(Vivek Vihar) PM2.5(DTU)

[mean, interval] [mean, interval] [mean, interval] [mean,interval]
Mar-Jun(2018) -- 104.59 10.47 89.55 7.39 85.98 8.05
Mar-Jun(2019) 99.38 7.41 97.62 8.48 77.65 5.59 84.81 6.13
Mar-Jun(2020) 64.31 5.44 49.11 6.45 52.78 4.08 57.94 4.74

4.3. Results: Modelling and Forecasting confidence interval using 30 experimental runs. Note that lower
We report the train and test performance (RMSE) mean and value of RMSE indicates better performance. We find that the
(±) 95% interval of the RMSE for different prediction horizons prediction RMSE in general increases across prediction hori-
for each model from 30 experimental runs. Table 6 shows the zons for all the models which is also natural as we use a speci-
mean and (±) 95% confidence interval for PM2.5 concentra- fied window size as input to predict the multi-step ahead values
tion for different monitoring stations for the months March to and the results deteriorate as the gap in the missing values in-
June in the years 2018, 2019 and 2020. We selected this par- creases with increasing prediction horizon. In Figure 6(b) and
ticular interval of months to show a comparison of PM2.5 con- Table 7, we find that the multivariate BD-LSTM performs the
centration in consecutive years for these COVID-19 lockdown best on train dataset generalizes well with test performance. We
months. We observe a significant decrease in PM2.5 concentra- find that ED-LSTM is the worst performer in this case with
tion in 2020 for all monitoring stations as compared to 2019 as a highest RMSE on both train and test dataset in comparison to
consequence of COVID-19 lockdown. In our first set of exper- other models. In Figure 6(a) and Table 7, we observe that mul-
iments, we compare the performance of different multivariate tivariate BD-LSTM generally has the best performance which
LSTM models (LSTM, ED-LSTM and BD-LSTM) where all significantly improved for step sizes 6-10 when compared to
the models are trained using Adam Optimizer using the dataset other models.
collected from the monitoring station at Anand Vihar. We also Since all the monitoring stations belong to Delhi region, we
show results with canonical feedforward neural network (FNN- use the model which performs the best on dataset from Anand
Adam) with Adam optimiser. We also show results for case Vihar to analyse the performance on the dataset from other sta-
of shuffling the training data (BD-LSTM*), univariate (UBD- tions namely, Bawana, DTU and Vivek Vihar. Taking into ac-
LSTM) and seasonal (SBD-LSTM) at Anand Vihar. Note that count the results for Anand Vihar, we consider BD-LSTM as
all LSTM models are multivariate except for UBD-LSTM. our best model, and further apply it other monitoring stations.
Figure 6(a) shows the prediction horizon RMSE for 10 step We also consider different strategies with BD-LSTM to train
ahead prediction, and Figure 6(b) shows the mean RMSE for and analyse the performance on dataset from different moni-
entire train and test datasets for Anand Vihar using different toring stations which includes shuffling the input window when
LSTM models. We show bar plots with mean and (±) 95% training BD-LSTM model (BD-LSTM*), univariate time-series
9
 analysis using BD-LSTM model (UBD-LSTM) and training the
BD-LSTM model using seasonal data (SBD-LSTM).
In Figure 7(a) and Table 7 results for Anand Vihar, we find
that BD-LSTM performs better on train dataset in comparison
to BD-LSTM* with almost similar performance on test dataset
while UBD-LSTM has highest RMSE on both train and test
dataset with highest confidence interval(error) of prediction.
In Figure 7(b) and Table 8 results for Bawana, we find that
BD-LSTM performs the best on train dataset in comparison
to BD-LSTM* and UBD-LSTM while BD-LSTM* performs
the best on test dataset amongst them. We find a similar trend
as Anand Vihar for UBD-LSTM with highest RMSE on both
train and test dataset and highest confidence interval of predic-
tion. In Figure 7(c) and Table 9 results for DTU, we find that
UBD-LSTM performs better than BD-LSTM and BD-LSTM*
on train dataset while BD-LSTM* performs the best on test
dataset among these three, we observe a similar trend for UBD-
LSTM with highest confidence of prediction on both train and
test dataset. In Figure 7(d) and Table 10 results for Vivek Vihar,
we find that BD-LSTM* performs the best on train dataset in
comparsion to BD-LSTM and UBD-LSTM, while BD-LSTM
performs the best on test dataset among them. UBD-LSTM has
the worst performance on both train and test dataset with largest
confidence interval for the predictions.
In the seasonal model for the monitoring stations (SBD-
(a) RMSE for different prediction horizons for Anand Vihar.
LSTM), we find least RMSE for both train and test dataset with
better performance on test dataset for Anand Vihar, Bawana
and DTU. For Vivek Vihar, we find that SBD-LSTM performs
the best on train dataset but it does not generalise well on test
dataset. We next evaluate the performance of SBD-LSTM in
the same plot for other models (BD-LSTM, BD-LSTM*,UBD-
LSTM). We cannot make a direct comparison between SBD-
LSTM and other models in terms of prediction accuracy, since
SBD-LSTM uses a subset of initial data (February-September,
2019) with lower range of PM2.5 values for these months as
shown in Figure 5.
Figure 8 provides a long-term (one month ahead) forecast
(11th December 2020 to 9th January 2021) of PM2.5 concentra-
tion values for different monitoring stations using UBD-LSTM
model. The plots provide mean and (±) 95% confidence inter-
val as uncertainty in predicted values for 30 experimental runs
for 8 hours interval for each day, covering a total of 720 hours
(b) RMSE for different models for Anand Vihar.
for the entire month. Figure 9 shows a comparison between ac-
tual and predicted values of PM2.5 concentration for September
Figure 6: Performance evaluation of respective methods for Anand Vihar (mean 2020 in terms of mean and (±) 95% confidence interval in pre-
RMSE with (±) 95% confidence interval of 30 experimental runs as uncertainty dicted values for 30 experimental runs for different monitoring
given by error bar). stations using SBD-LSTM model.

5. Discussion
In the respective experiments, we evaluated the performance
of different models on the time-series problem and studied the
effect of COVID-19 lockdown on prediction performance of
different models for different seasons. We find that BD-LSTM
is our best performer in terms of both train and test performance
with better generalization when compared to other LSTM based
models for Anand Vihar. This could be due to the BD-LSTM
10
 FNN-Adam LSTM ED-LSTM BD-LSTM BD-LSTM* UBD-LSTM SBD-LSTM
Train 1939.38± 30.79 2017.92± 33.34 2101.77± 34.16 1797.43± 32.68 1840.42± 28.45 2178.68± 76.78 1545.79± 34.30
Test 1821.29± 37.45 1797.96± 37.64 1871.46± 37.17 1808.26± 34.04 1810.86± 36.21 1822.95± 65.80 1477.35± 39.34
Step-1 1177.20± 21.02 1152.12± 19.25 1195.88± 26.43 1153.02± 22.00 1130.93± 19.09 317.40± 28.01 843.21± 27.57
Step-2 1397.96± 32.76 1390.28± 26.98 1512.29± 36.34 1494.42± 39.38 1504.22± 50.99 1451.15± 58.28 1034.96± 25.71
Step-3 1620.34± 37.52 1615.66± 39.30 1695.36± 37.67 1731.97± 49.72 1718.83± 56.19 1690.91± 71.69 1283.02± 51.79
Step-4 1782.44± 36.65 1713.78± 32.38 1810.73± 41.03 1776.68± 47.31 1736.79± 40.89 1863.03± 67.08 1425.18± 68.46
Step-5 1873.96± 42.76 1812.06± 31.15 1925.16± 41.85 1815.68± 25.61 1806.75± 46.08 2010.27± 62.85 1508.44± 67.65
Step-6 1956.12± 38.44 1923.46± 39.53 2000.59± 46.58 1894.14± 25.17 1894.19± 42.10 2205.94± 63.72 1675.28± 73.59
Step-7 2053.13± 41.38 2001.73± 41.10 2079.05± 46.45 1971.98± 32.10 1983.71± 43.13 2223.07± 65.43 1685.23± 56.25
Step 8 2064.14± 41.96 2069.54± 36.31 2128.60± 45.86 2020.71± 39.29 2054.23± 45.53 2174.59± 44.03 1718.71± 46.06
Step 9 2140.37± 46.04 2146.61± 45.23 2170.13± 45.64 2093.84± 35.45 2149.13± 37.42 2174.37± 59.44 1827.85± 48.45
Step 10 2147.23± 43.46 2154.41± 44.34 2196.85± 43.78 2130.20± 40.49 2129.89± 32.62 2118.77± 65.65 1771.62± 41.05

Table 7: Performance by RMSE (mean) and 95 % confidence interval(±) of 30 experimental runs for Anand Vihar.

BD-LSTM BD-LSTM* UBD-LSTM SBD-LSTM

Train 1949.35±31.17 2027.01±33.72 2081.15± 74.34 1708.78± 35.58
Test 1833.16±43.64 1732.85±40.63 1842.50± 71.36 1281.22± 35.57
Step-1 1100.56±26.64 1017.65±22.34 297.75± 25.75 683.00±29.18
Step-2 1385.24±37.23 1253.59±27.32 1297.76± 36.08 937.01±23.71
Step-3 1651.61±41.44 1534.16±24.59 1591.85± 40.72 1095.53± 36.70
Step-4 1721.90±37.51 1659.84±29.50 1834.38± 38.84 1173.04± 33.13
Step-5 1875.45±38.82 1835.25±25.71 2163.71± 52.14 1249.93± 38.67
Step-6 1919.22±45.39 1880.20±26.49 2323.89±63.87 1384.16± 37.51
Step-7 2027.93±60.23 1941.07±32.05 2479.18± 71.26 1465.01± 34.17
Step 8 2160.84±67.78 2040.73±42.58 2252.38± 61.94 1538.47± 35.68
Step 9 2235.11±72.46 2075.96±44.83 2112.08±49.15 1618.90± 32.90
Step 10 2253.73±70.02 2090.06±47.47 2072.01±43.98 1667.10±37.50

Table 8: Performance by RMSE (mean) and 95 % confidence interval(±) of 30 experimental runs using different BD-LSTM models for Bawana.

BD-LSTM BD-LSTM* UBD-LSTM SBD-LSTM

Train 1894.97±35.20 1867.04±34.88 1837.32±63.35 1394.28± 25.60
Test 1707.09±34.12 1655.13±34.73 1871.76± 68.03 1346.61±41.85
Step-1 1082.65±35.59 990.05±24.93 376.30± 46.39 661.50±26.35
Step-2 1412.94±37.73 1312.39±28.11 1368.83± 51.53 1009.89±22.47
Step-3 1595.74± 42.79 1541.37±36.48 1890.12± 98.02 1164.77± 26.02
Step-4 1702.60±51.73 1616.10±36.87 2030.47± 92.59 1165.97± 26.92
Step-5 1804.52±52.03 1728.34±42.14 2224.31± 92.41 1414.22± 42.56
Step-6 1841.11±57.09 1788.62±41.66 2272.49± 102.77 1462.59± 45.03
Step-7 1894.75±75.12 1816.18±38.03 2236.76± 90.09 1411.73± 54.63
Step 8 1946.97± 70.98 1929.45±41.18 2143.55±67.55 1743.37± 83.88
Step 9 1910.16±56.44 1953.05±36.84 2165.78± 79.03 1786.94±94.11
Step 10 1879.49±47.86 1875.75±37.62 2008.97± 57.63 1645.09±86.99

Table 9: RMSE (mean) and 95 % confidence interval(±) of 30 experimental runs using different BD-LSTM models for DTU.

11
 (a) RMSE(mean) with error bar for Anand Vihar. (b) RMSE(mean) with error bar for Bawana.

(c) RMSE(mean) with error bar for DTU. (d) RMSE(mean) with error bar for Vivek Vihar.

Figure 7: Performance by RMSE (mean) with (±) 95% confidence interval as error bar across 10 prediction horizons of 30 experimental runs using respective
BD-LSTM models. We show results for train and test datasets collected from different monitoring stations in Delhi.

architecture which features both forward and backward infor- ate models which uses a set of features as input that helps with
mation for a sequence at each time-step using two LSTM layers additional information for predicting future steps. The seasonal
[95]. We find that BD-LSTM performs the best for larger time model (SBD-LSTM) performs slightly better on test dataset for
steps (6-10) with better generalization as compared to other Anand Vihar, Bawana and DTU along with good generaliza-
models which indicates that BD-LSTM captures better infor- tion performance which indicates that the model is robust to the
mation for long-term ahead prediction. This could be due to effects of COVID-19 lockdown period. It is important to eval-
the combination of memory cells with two LSTM layers which uate the effect on performance of different models due to sea-
are able to capture salient features in a temporal sequence when sonal decrease of PM2.5 values during COVID-19 lockdown
compared to canonical LSTM models. Although BD-LSTM period when compared to previous years. We find that different
based models have been mainly used for different language LSTM models with different training strategies are quite robust
modelling, we find them promising for long-term time-series in modeling the effects of COVOD-19 lockdown for both dur-
prediction problems. ing and partial lockdown periods.

In the respective experiments, we also found that BD-LSTM Recent focus has been on studying effects of the lock downs
and BD-LSTM* perform well in predicting future PM2.5 val- imposed in different cities of India on the spatial patterns of air
ues for different monitoring stations which indicates that they quality both during pre-lockdown and during-lockdown phases.
are not affected much due to COVID-19 lockdown phase. The results have demonstrated a direct implication towards the
UBD-LSTM has the highest RMSE and confidence interval for decrease in the concentration of selected pollutants such as
test datasets which indicates that it is not robust. The poor per- PM2.5 and PM10, with maximum decrease(more than 50%)
formance which could be due to the fact that UBD-LSTM uses in comparison to the pre-lockdown phase [104]. There has
only one features to predict future steps in contrast to multivari- also been a significant decrease in the concentration of Nitro-
12
 BD-LSTM BD-LSTM* UBD-LSTM SBD-LSTM
Train 1820.40±41.58 1779.95±33.98 1994.93± 69.00 1434.41±31.50
Test 1643.59±40.90 1696.24±42.56 1742.39±59.96 1706.87± 55.33
Step-1 782.12± 14.33 792.82± 12.90 302.54±23.65 775.95±21.48
Step-2 1352.21±28.23 1392.97±24.17 1461.58± 34.56 1139.47± 32.01
Step-3 1548.57±32.00 1612.39±34.50 1640.58± 37.38 1398.11± 54.94
Step-4 1618.60±26.98 1643.44±31.46 1847.20± 53.83 1493.66± 47.42
Step-5 1689.86±31.89 1764.74±31.53 2012.07±57.44 1785.54± 53.64
Step-6 1752.82±27.40 1844.85±43.18 2038.00± 47.06 1936.81± 61.08
Step-7 1775.63±40.77 1881.67±50.28 2054.06±41.14 2013.61± 62.70
Step 8 1920.66±53.87 2002.18±56.76 2058.89± 41.50 2180.80±66.23
Step 9 2010.76±52.36 2061.93±56.88 2019.43± 34.69 2251.07± 68.98
Step 10 1984.70±47.81 1965.43±50.60 1989.56± 40.73 2093.66± 63.69

Table 10: RMSE (mean) and 95 % confidence interval(±) of 30 experimental runs using different BD-LSTM models for Vivek Vihar.

gen Dioxide(NO2) which has been a major pollutant in reputed in relation to forest fires around the world.
cities of India such as Delhi and Mumbai [105]. There is a
positive correlation between PM2.5 concentration and COVID-
19 for countries such as India and Pakistan [106]. Apart from 6. Conclusion and Future Work
the major air pollutants, there has been a significant decrease
In this paper, we applied deep learning via different LSTM
in aerosol optical depth (AOD) and lightning activities in many
models for univariate and multivariate modelling for short-term
urban and mining regions of India [107, 108].
and long term air quality prediction taking into account the four
There is a large effort by the Indian government to move to- base-stations from Delhi, India. Although we selected a subset
wards clean energy with a massive investment in solar and wind of the base stations, the methodology can be applied to rest of
energy which can have benefits [109, 110]. Moreover, India the base stations in Delhi or in other parts of the world. Our re-
also has plans to go fully electric in the transportation sector by sults show that multivariate bi-directional LSTM model shows
2030 [111]. Currently, India has began manufacturing electric best performance, and rest of the LSTM models have certain
vehicles with plans to have fully electric rail network by 2024 strengths and limitations which need to be evaluated prior to
[112]. Renewable energy is expected to rise from 27 % of total developing a system that provides rigours uncertainty quantifi-
energy demand in 2014 to around 43 % by 2040 [113]. In many cation in predictions. We also found that COVID-19 had a
developing countries, it is clear that an increase of population significant effect on the air quality during full lockdown im-
density increases the air pollution; however, this can change plemented for few months and afterwards, there was unprece-
with type of energy used. In a recent study, in the case of China, dented growth of poor air quality which has a seasonal effect as
it has been shown that the increase in population density will compared to previous years. We provide open source software
reduces air pollution which has been due to clean energy and framework and open data which can be used for further verifica-
public transportation [114]. A study on the effects of meteoro- tion and also application for studying or developing prediction
logical conditions and air pollution on COVID-19 transmission models in other places.
from 219 Chinese cities found that air pollution indicators were
positively correlated with new confirmed cases and increase in
cases were linked with air-quality index [115]. Such study in 7. Software and Data
case of India would be needed and prediction of air-quality in-
dicators could help in preventative measures. We provide Python based open source implementation along
with the data for different experiments and design of respective
In future work, Bayesian deep learning methods can be
method for further research 3 .
used to provide robust uncertainty quantification in predictions
which can extend Bayesian neural networks used for time series
prediction [116]. Other learning strategies such as multi-task References
and transfer learning in conjunction with Bayesian inference
can be used to develop improved models [117, 118] which can [1] H. R. Max Roser and E. Ortiz-Ospina, “World population growth,”
Our World in Data, 2013, https://ourworldindata.org/world-population-
take into account existing deep learning models for COVID-19 growth.
infections in India [119]. Moreover, we envision the application [2] U. Desa, “United nations department of economic and social affairs,
of the proposed framework in other parts of India and rest of the population division. world population prospects: The 2015 revision, key
world which has decline in air quality. A web-based framework findings and advance tables,” in Technical Report: Working Paper No.
ESA/P/WP. 241, 2015.
can be implemented by the respective authorities that can be
used to provide proper weekly and monthly planning, which
could involve the way traffic is managed for different seasons. 3 https://github.com/sydney-machine-learning/

The methodology can also be extended for air quality prediction airpollution_deeplearning

13
 (a) PM2.5 Concentration prediction for next one month for Anand Vihar. (b) PM2.5 Concentration prediction for next one month for Bawana.

(c) PM2.5 Concentration prediction for next one month for DTU. (d) PM2.5 Concentration prediction for next one month for Vivek Vihar.

Figure 8: PM2.5 Concentration prediction with Mean and (±) 95% confidence interval as error for next one month using UBD-LSTM model for different monitoring
stations.

[3] P. J. Taylor and B. Derudder, World city network: a global urban analy- contribution of outdoor air pollution sources to premature mortality on a
sis. Routledge, 2004. global scale,” Nature, vol. 525, no. 7569, pp. 367–371, 2015.
[4] WHO, “Air pollution levels rising in many of the world’s poorest cities,” [12] “Unique identification authority of India,” May 2020, [Online;
2016. accessed 16-July-2020]. [Online]. Available: https://uidai.gov.in/
[5] UNEP, “Status of fuel quality and vehicle emission standards latin amer- images/state-wise-aadhaar-saturation.pdf
ica,” 2016. [13] “6 of the world’s 10 most polluted cities are in In-
[6] S. L. Harlan and D. M. Ruddell, “Climate change and health in cities: dia,” August 2020, [Online; accessed 22-August-2020].
impacts of heat and air pollution and potential co-benefits from mitiga- [Online]. Available: https://www.weforum.org/agenda/2020/03/
tion and adaptation,” Current Opinion in Environmental Sustainability, 6-of-the-world-s-10-most-polluted-cities-are-in-india/
vol. 3, no. 3, pp. 126–134, 2011. [14] “7 million premature deaths annually linked to air pollution,”
[7] P. Gupta, S. A. Christopher, J. Wang, R. Gehrig, Y. Lee, and N. Kumar, August 2020, [Online; accessed 22-August-2020]. [Online]. Available:
“Satellite remote sensing of particulate matter and air quality assessment https://www.who.int/mediacentre/news/releases/2014/air-pollution/en/
over global cities,” Atmospheric Environment, vol. 40, no. 30, pp. 5880– [15] J. S. Apte, M. Brauer, A. J. Cohen, M. Ezzati, and C. A. Pope III, “Ambi-
5892, 2006. ent pm2. 5 reduces global and regional life expectancy,” Environmental
[8] Y.-L. Zhang and F. Cao, “Fine particulate matter (pm 2.5) in china at a Science & Technology Letters, vol. 5, no. 9, pp. 546–551, 2018.
city level,” Scientific reports, vol. 5, p. 14884, 2015. [16] D. H. Cusworth, L. J. Mickley, M. P. Sulprizio, T. Liu, M. E. Marlier,
[9] C. Liu, P.-C. Hsu, H.-W. Lee, M. Ye, G. Zheng, N. Liu, W. Li, and R. S. DeFries, S. K. Guttikunda, and P. Gupta, “Quantifying the influ-
Y. Cui, “Transparent air filter for high-efficiency pm 2.5 capture,” Nature ence of agricultural fires in northwest india on urban air pollution in
communications, vol. 6, no. 1, pp. 1–9, 2015. delhi, india,” Environmental Research Letters, vol. 13, no. 4, p. 044018,
[10] M. Franklin, A. Zeka, and J. Schwartz, “Association between pm 2.5 and 2018.
all-cause and specific-cause mortality in 27 us communities,” Journal of [17] W. Hogland and J. Stenis, “Assessment and system analysis of industrial
exposure science & environmental epidemiology, vol. 17, no. 3, pp. 279– waste management,” Waste Management, vol. 20, no. 7, pp. 537–543,
287, 2007. 2000.
[11] J. Lelieveld, J. S. Evans, M. Fnais, D. Giannadaki, and A. Pozzer, “The [18] V.-D. Le, T.-C. Bui, and S.-K. Cha, “Spatiotemporal deep learning

14
 (a) PM2.5 Concentration(actual and predicted) for Anand Vihar. (b) PM2.5 Concentration(actual and predicted) for Bawana.

(c) PM2.5 Concentration(actual and predicted) for DTU. (d) PM2.5 Concentration(actual and predicted) for Vivek Vihar.

Figure 9: PM2.5 Concentration (actual and predicted) with Mean and (±) 95% confidence interval as error of the predicted values for September, 2020 using
SBD-LSTM model for different monitoring stations.

model for citywide air pollution interpolation and prediction,” in 2020 Machinery, 2008, p. 160–167. [Online]. Available: https://doi.org/10.
IEEE International Conference on Big Data and Smart Computing (Big- 1145/1390156.1390177
Comp). IEEE, 2020, pp. 55–62. [26] J. Fan, Y. Gao, and H. Luo, “Integrating concept ontology and multitask
[19] X. Li, L. Peng, Y. Hu, J. Shao, and T. Chi, “Deep learning architec- learning to achieve more effective classifier training for multilevel image
ture for air quality predictions,” Environmental Science and Pollution annotation,” IEEE transactions on image processing : a publication of
Research, vol. 23, no. 22, pp. 22 408–22 417, 2016. the IEEE Signal Processing Society, vol. 17, pp. 407–26, 04 2008.
[20] T.-C. Bui, V.-D. Le, and S.-K. Cha, “A deep learning approach for [27] C. Widmer, J. Leiva, Y. Altun, and G. Rätsch, “Leveraging sequence
forecasting air pollution in south korea using lstm,” arXiv preprint classification by taxonomy-based multitask learning,” vol. 6044, 04
arXiv:1804.07891, 2018. 2010, pp. 522–534.
[21] G. Grell, S. Peckham, R. Schmitz, S. McKeen, G. Frost, W. Skamarock, [28] A. Lindbeck and D. J. Snower, “Multitask learning and the reorganiza-
and B. Eder, “Fully coupled “online” chemistry in the wrf model,” At- tion of work: from tayloristic to holistic organization,” Journal of labor
mospheric Environment, vol. 39, pp. 6957–6975, 12 2005. economics, vol. 18, no. 3, pp. 353–376, 2000.
[22] L. Emmons, S. Walters, P. Hess, J.-F. Lamarque, G. Pfister, F. D, [29] C. J. Patel, “Analytic complexity and challenges in identifying mixtures
C. Granier, A. Guenther, K. D, T. Laepple, O. J, X. Tie, T. G, C. Wied- of exposures associated with phenotypes in the exposome era,” Current
inmyer, S. Baughcum, and S. Kloster, “Description and evaluation of epidemiology reports, vol. 4, no. 1, pp. 22–30, 2017.
the model for ozone and related chemical tracers, version 4 (mozart-4),” [30] J. Schmidhuber, “Deep learning in neural networks: An overview,” Neu-
Geoscientific Model Development Discussions, vol. 3, 01 2009. ral networks, vol. 61, pp. 85–117, 2015.
[23] H. Karimian, Q. Li, C. LI, L. Jin, J. Fan, and Y. Li, “An improved method [31] J. L. Elman and D. Zipser, “Learning the hidden structure of speech,”
for monitoring fine particulate matter mass concentrations via satellite The Journal of the Acoustical Society of America, vol. 83, no. 4, pp.
remote sensing,” Aerosol and Air Quality Research, vol. 16, pp. 1081– 1615–1626, 1988. [Online]. Available: http://link.aip.org/link/?JAS/83/
1092, 01 2016. 1615/1
[24] Z. Yuan, X. Zhou, T. Yang, and J. Tamerius, “Predicting traffic accidents [32] P. J. Werbos, “Backpropagation through time: what it does and how to
through heterogeneous urban data : A case study,” 2017. do it,” Proceedings of the IEEE, vol. 78, no. 10, pp. 1550–1560, 1990.
[25] R. Collobert and J. Weston, “A unified architecture for natural [33] S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural
language processing: Deep neural networks with multitask learning,” in computation, vol. 9, no. 8, pp. 1735–1780, 1997.
Proceedings of the 25th International Conference on Machine Learning, [34] R. Chandra, “Competition and collaboration in cooperative coevolution
ser. ICML ’08. New York, NY, USA: Association for Computing of Elman recurrent neural networks for time-series prediction,” Neural

15
 Networks and Learning Systems, IEEE Transactions on, vol. 26, pp. Environment, p. 139282, 2020.
3123–3136, 2015. [55] F. Dutheil, J. S. Baker, and V. Navel, “COVID-19 as a factor influencing
[35] S. Hochreiter, “The vanishing gradient problem during learning recur- air pollution?” Environmental Pollution (Barking, Essex: 1987), vol.
rent neural nets and problem solutions,” International Journal of Un- 263, p. 114466, 2020.
certainty, Fuzziness and Knowledge-Based Systems, vol. 6, no. 02, pp. [56] R. Tosepu, J. Gunawan, D. S. Effendy, H. Lestari, H. Bahar, P. As-
107–116, 1998. fian et al., “Correlation between weather and COVID-19 pandemic in
[36] Y. Bengio, P. Simard, and P. Frasconi, “Learning long-term dependen- jakarta, indonesia,” Science of The Total Environment, p. 138436, 2020.
cies with gradient descent is difficult,” Neural Networks, IEEE Transac- [57] Y. Ma, Y. Zhao, J. Liu, X. He, B. Wang, S. Fu, J. Yan, J. Niu, J. Zhou,
tions on, vol. 5, no. 2, pp. 157–166, 1994. and B. Luo, “Effects of temperature variation and humidity on the death
[37] J. Chung, C. Gulcehre, K. Cho, and Y. Bengio, “Empirical evaluation of of COVID-19 in Wuhan, China,” Science of The Total Environment, p.
gated recurrent neural networks on sequence modeling,” arXiv preprint 138226, 2020.
arXiv:1412.3555, 2014. [58] W. Wang, Z. Xu, and W.-Z. Lu, “Three improved neural network models
[38] K. Cho, B. Van Merriënboer, C. Gulcehre, D. Bahdanau, F. Bougares, for air quality forecasting,” Engineering Computations, vol. 20, pp. 192–
H. Schwenk, and Y. Bengio, “Learning phrase representations using 210, 03 2003.
rnn encoder-decoder for statistical machine translation,” arXiv preprint [59] G. Corani, “Air quality prediction in milan: Feed-forward neural net-
arXiv:1406.1078, 2014. works, pruned neural networks and lazy learning,” Ecological Mod-
[39] M. Schuster and K. K. Paliwal, “Bidirectional recurrent neural net- elling, pp. 513–529, 07 2005.
works,” IEEE Transactions on Signal Processing, vol. 45, no. 11, pp. [60] H. Peng, “Air quality prediction by machine learning methods,” 2015.
2673–2681, 1997. [61] M. Fu, W. Wang, Z. Le, and M. Safaei khorram, “Prediction of par-
[40] A. Graves and J. Schmidhuber, “2005 special issue: Framewise ticular matter concentrations by developed feed-forward neural network
phoneme classification with bidirectional lstm and other neural network with rolling mechanism and gray model,” Neural Computing and Appli-
architectures,” Neural Netw., vol. 18, no. 5–6, p. 602–610, Jun. 2005. cations, 02 2015.
[Online]. Available: https://doi.org/10.1016/j.neunet.2005.06.042 [62] Y.-S. Chang, H.-T. Chiao, S. Abimannan, Y.-P. Huang, Y.-T. Tsai,
[41] A. E. Gorbalenya, S. C. Baker, R. S. Baric, R. J. de Groot, C. Drosten, and K.-M. Lin, “An lstm-based aggregated model for air pollution
A. A. Gulyaeva, B. L. Haagmans, C. Lauber, A. M. Leontovich, B. W. forecasting,” Atmospheric Pollution Research, vol. 11, no. 8, pp. 1451 –
Neuman, D. Penzar, L. L. M. P. Stanley Perlman10, D. V. Samborskiy, 1463, 2020. [Online]. Available: http://www.sciencedirect.com/science/
I. A. Sidorov, I. Sola, and J. Ziebuhr, “The species severe acute respira- article/pii/S1309104220301215
tory syndrome-related coronavirus: classifying 2019-ncov and naming [63] D. Xiao, J. Zheng, C. Pain, and I. Navon, “Machine learning-based rapid
it sars-cov-2,” Nature Microbiology, vol. 5, no. 4, p. 536, 2020. response tools for regional air pollution modelling,” Atmospheric Envi-
[42] V. Monteil, H. Kwon, P. Prado, A. Hagelkrüys, R. A. Wimmer, M. Stahl, ronment, vol. 199, 11 2018.
A. Leopoldi, E. Garreta, C. H. Del Pozo, F. Prosper et al., “Inhibition of [64] D. Zhu, C. Cai, T. Yang, and X. Zhou, “A machine learning approach for
sars-cov-2 infections in engineered human tissues using clinical-grade air quality prediction: Model regularization and optimization,” Big Data
soluble human ace2,” Cell, 2020. and Cognitive Computing, vol. 2, p. 5, 02 2018.
[43] W. H. Organization et al., “Coronavirus disease 2019 ( COVID-19): sit- [65] E. Kim, H. Park, Y.-C. Hong, M. Ha, Y. Kim, B.-N. Kim, Y. Kim, Y.-
uation report, 72,” 2020. M. Roh, B.-E. Lee, J.-M. Ryu et al., “Prenatal exposure to pm10 and
[44] D. Cucinotta and M. Vanelli, “WHO declares COVID-19 a pandemic,” no2 and children’s neurodevelopment from birth to 24 months of age:
Acta bio-medica: Atenei Parmensis, vol. 91, no. 1, pp. 157–160, 2020. Mothers and children’s environmental health (moceh) study,” Science of
[45] A. Atkeson, “What will be the economic impact of COVID-19 in the the Total Environment, vol. 481, pp. 439–445, 2014.
US? rough estimates of disease scenarios,” National Bureau of Eco- [66] Y. Huang, Y. Wu, and W. Zhang, “Comprehensive identification and
nomic Research, Tech. Rep., 2020. isolation policies have effectively suppressed the spread of COVID-19,”
[46] N. Fernandes, “Economic effects of coronavirus outbreak (covid-19) on Chaos, Solitons& Fractals, vol. 139, p. 110041, 2020.
the world economy,” Available at SSRN 3557504, 2020. [67] X. Ma, I. Longley, J. Gao, and J. Salmond, “Assessing schoolchildren’s
[47] M. Maliszewska, A. Mattoo, and D. Van Der Mensbrugghe, “The poten- exposure to air pollution during the daily commute - a systematic re-
tial impact of COVID-19 on gdp and trade: A preliminary assessment,” view,” Science of The Total Environment, vol. 737, p. 140389, 2020.
2020. [68] K. Liu, B.-Y. Yang, Y. Guo, M. S. Bloom, S. C. Dharmage, L. D.
[48] A. Kerimray, N. Baimatova, O. P. Ibragimova, B. Bukenov, B. Kenessov, Knibbs, J. Heinrich, A. Leskinen, S. Lin, L. Morawska, B. Jalaludin,
P. Plotitsyn, and F. Karaca, “Assessing air quality changes in large cities I. Markevych, P. Jalava, M. Komppula, Y. Yu, M. Gao, Y. Zhou, H.-
during COVID-19 lockdowns: The impacts of traffic-free urban con- Y. Yu, L.-W. Hu, X.-W. Zeng, and G.-H. Dong, “The role of influenza
ditions in almaty, kazakhstan,” Science of the Total Environment, p. vaccination in mitigating the adverse impact of ambient air pollution
139179, 2020. on lung function in children: New insights from the seven northeast-
[49] G. Dantas, B. Siciliano, B. B. França, C. M. da Silva, and G. Arbilla, ern cities study in china,” Environmental Research, vol. 187, p. 109624,
“The impact of COVID-19 partial lockdown on the air quality of the 2020.
city of rio de janeiro, brazil,” Science of the Total Environment, vol. 729, [69] Z. Wang, W. Wei, and F. Zheng, “Effects of industrial air pollution on the
p. 139085, 2020. technical efficiency of agricultural production: Evidence from china,”
[50] Z. Yongjian, X. Jingu, H. Fengming, and C. Liqing, “Association be- Environmental Impact Assessment Review, vol. 83, p. 106407, 2020.
tween short-term exposure to air pollution and COVID-19 infection: Ev- [70] M. Xu, Y. Wang, and Y. Tu, “Uncovering the invisible effect of air pol-
idence from china,” Science of the total environment, p. 138704, 2020. lution on stock returns: A moderation and mediation analysis,” Finance
[51] P. Wang, K. Chen, S. Zhu, P. Wang, and H. Zhang, “Severe air pol- Research Letters, p. 101646, 2020.
lution events not avoided by reduced anthropogenic activities during [71] “The worlds cities in 2016: data booklet,” June 2016, [United
COVID-19 outbreak,” Resources, Conservation and Recycling, vol. 158, Nations, Online; accessed 24-August-2020]. [Online]. Available: https:
p. 104814, 2020. //www.un.org/en/development/desa/population/publications/index.asp
[52] “India coronavirus: Can the covid-19 lockdown spark a clean air move- [72] “Usual suspects: Vehicles, industrial emissions behind foul
ment?” August 2020, [Online; accessed 24-August-2020]. [Online]. play,” August 2018, [Online; accessed 24-August-2020].
Available: https://www.bbc.com/news/world-asia-india-52313972 [Online]. Available: https://timesofindia.indiatimes.com/city/delhi/
[53] R. Bao and A. Zhang, “Does lockdown reduce air pollution? evidence usual-suspects-vehicles-industrial-emissions-behind-foul-play-all-year/
from 44 cities in northern China,” Science of The Total Environment, vol. articleshow/66228517.cms
731, p. 139052, 2020. [73] “Delhi breathed easier from january to april,”
[54] L. Li, Q. Li, L. Huang, Q. Wang, A. Zhu, J. Xu, Z. Liu, H. Li, L. Shi, June 2017, [Online; accessed 24-August-2020]. [On-
R. Li et al., “Air quality changes during the COVID-19 lockdown over line]. Available: https://timesofindia.indiatimes.com/city/delhi/
the yangtze river delta region: An insight into the impact of human ac- delhi-breathed-easier-from-january-to-april/articleshow/59011204.cms
tivity pattern changes on air pollution variation,” Science of The Total [74] “Air pollution: Delhi enjoys cleanest february in three years,”

16
 February 2018, [Online; accessed 24-August-2020]. [On- works : the official journal of the International Neural Network Society,
line]. Available: https://www.hindustantimes.com/delhi-news/ vol. 18, pp. 602–10, 07 2005.
air-pollution-delhi-enjoys-cleanest-february-in-three-years/ [96] Y. Fan, Y. Qian, F.-L. Xie, and F. K. Soong, “Tts synthesis with bidirec-
story-SANlmslHev8ifFgZbh3WXI.html tional lstm based recurrent neural networks,” in INTERSPEECH, 2014.
[75] R. Suliankatchi, B. Nongkynrih, and S. Gupta, “Air pollution in delhi: [97] A. Graves, N. Jaitly, and A.-r. Mohamed, “Hybrid speech recognition
Its magnitude and effects on health,” Indian Journal of Community with deep bidirectional lstm,” in 2013 IEEE workshop on automatic
Medicine, vol. 38, pp. 4–8, 01 2013. speech recognition and understanding. IEEE, 2013, pp. 273–278.
[76] A. Comrie, “A synoptic climatology of rural ozone pollution at three [98] Z. Ding, R. Xia, J. Yu, X. Li, and J. Yang, “Densely connected bidi-
forest sites in pennsylvania,” Atmospheric Environment, vol. 28, no. 9, rectional lstm with applications to sentence classification,” in CCF In-
pp. 1601–1614, May 1994. ternational Conference on Natural Language Processing and Chinese
[77] A. Degaetano and O. Doherty, “Temporal, spatial and meteorological Computing. Springer, 2018, pp. 278–287.
variations in hourly pm 2.5 concentration extremes in new york city,” [99] I. Sutskever, O. Vinyals, and Q. V. Le, “Sequence to sequence
Atmospheric Environment - ATMOS ENVIRON, vol. 38, pp. 1547–1558, learning with neural networks,” in Advances in Neural Information
04 2004. Processing Systems 27, Z. Ghahramani, M. Welling, C. Cortes, N. D.
[78] B. K. Eder, J. M. Davis, and P. Bloomfield, “An Automated Lawrence, and K. Q. Weinberger, Eds. Curran Associates, Inc.,
Classification Scheme Designed to Better Elucidate the Dependence 2014, pp. 3104–3112. [Online]. Available: http://papers.nips.cc/paper/
of Ozone on Meteorology,” Journal of Applied Meteorology, vol. 33, 5346-sequence-to-sequence-learning-with-neural-networks.pdf
no. 10, pp. 1182–1199, 10 1994. [Online]. Available: https: [100] T. Wang, P. Chen, K. Amaral, and J. Qiang, “An experimental study
//doi.org/10.1175/1520-0450(1994)033h1182:AACSDTi2.0.CO;2 of lstm encoder-decoder model for text simplification,” arXiv preprint
[79] M. P. Zelenka, “An analysis of the meteorological parameters affecting arXiv:1609.03663, 2016.
ambient concentrations of acid aerosols in Uniontown, Pennsylvania,” [101] A. Zeyer, P. Bahar, K. Irie, R. Schlüter, and H. Ney, “A comparison of
Atmospheric Environment, vol. 31, no. 6, pp. 869–878, Jan. 1997. transformer and lstm encoder decoder models for asr,” in 2019 IEEE
[80] A. Y. Watson, R. R. Bates, and D. Kennedy, “Atmospheric transport and Automatic Speech Recognition and Understanding Workshop (ASRU).
dispersion of air pollutants associated with vehicular emissions,” in Air IEEE, 2019, pp. 8–15.
Pollution, the Automobile, and Public Health. National Academies [102] K. Yao and G. Zweig, “Sequence-to-sequence neural net models for
Press (US), 1988. grapheme-to-phoneme conversion,” arXiv preprint arXiv:1506.00196,
[81] L. Natsagdorj, D. Jugder, and Y. Chung, “Analysis of dust storms 2015.
observed in mongolia during 1937–1999,” Atmospheric Environment, [103] S. Masum, Y. Liu, and J. Chiverton, Multi-step Time Series Forecasting
vol. 37, pp. 1401–1411, 03 2003. of Electric Load Using Machine Learning Models, 01 2018, pp. 148–
[82] R. Zalakeviciute, J. López-Villada, and Y. Rybarczyk, “Contrasted ef- 159.
fects of relative humidity and precipitation on urban pm2. 5 pollution in [104] S. Mahato, S. Pal, and K. G. Ghosh, “Effect of lockdown amid covid-
high elevation urban areas,” Sustainability, vol. 10, no. 6, p. 2064, 2018. 19 pandemic on air quality of the megacity delhi, india,” Science of the
[83] S. Seinfeld, J.H.; Pandis, “Atmospheric chemistry and physics: From air Total Environment, p. 139086, 2020.
pollution to climate change,” 2016. [105] K. Shehzad, M. Sarfraz, and S. G. M. Shah, “The impact of covid-19 as
[84] B. R. Appel, Y. Tokiwa, J. Hsu, E. L. Kothny, and E. Hahn, “Visibility as a necessary evil on air pollution in india during the lockdown,” Environ-
related to atmospheric aerosol constituents,” Atmospheric Environment, mental Pollution, vol. 266, p. 115080, 2020.
vol. 19, no. 9, pp. 1525–1534, Jan. 1985. [106] S. Sipra, M. M. Abrar, M. Iqbal, E. Haider, and H. M. H. Shoukat, “Can
[85] S. Twomey, “The Influence of Pollution on the Shortwave Albedo pm2. 5 pollution worsen the death rate due to covid-19 in india and pak-
of Clouds,” Journal of the Atmospheric Sciences, vol. 34, no. 7, pp. istan?” The Science of the Total Environment, vol. 742, p. 140557, 2020.
1149–1152, 07 1977. [Online]. Available: https://doi.org/10.1175/ [107] A. K. Ranjan, A. Patra, and A. Gorai, “Effect of lockdown due to sars
1520-0469(1977)034h1149:TIOPOTi2.0.CO;2 covid-19 on aerosol optical depth (aod) over urban and mining regions
[86] H. Akbari, “Shade trees reduce building energy use and co2 emissions in india,” Science of the Total Environment, vol. 745, p. 141024, 2020.
from power plants,” Environmental pollution (Barking, Essex : 1987), [108] I. Chowdhuri, S. C. Pal, A. Saha, R. Chakrabortty, M. Ghosh, and P. Roy,
vol. 116 Suppl 1, pp. S119–26, 02 2002. “Significant decrease of lightning activities during covid-19 lockdown
[87] G. o. I. Central Pollution Control Board, CCR, accessed 20 period over kolkata megacity in india,” Science of the Total Environment,
December, 2020. [Online]. Available: https://app.cpcbccr.com/ccr/#/ vol. 747, p. 141321, 2020.
caaqm-dashboard/caaqm-landing [109] M. Goel, “Solar rooftop in india: Policies, challenges and outlook,”
[88] J. L. Elman, “Finding structure in time,” Cognitive Science, vol. 14, pp. Green Energy & Environment, vol. 1, no. 2, pp. 129–137, 2016.
179–211, 1990. [110] P. K. Chaurasiya, V. Warudkar, and S. Ahmed, “Wind energy develop-
[89] C. W. Omlin, K. K. Thornber, and C. L. Giles, “Fuzzy finite state au- ment and policy in india: A review,” Energy Strategy Reviews, vol. 24,
tomata can be deterministically encoded into recurrent neural networks,” pp. 342–357, 2019.
IEEE Trans. Fuzzy Syst., vol. 6, pp. 76–89, 1998. [111] R. Vidhi and P. Shrivastava, “A review of electric vehicle lifecycle emis-
[90] C. W. Omlin and C. L. Giles, “Training second-order recurrent neural sions and policy recommendations to increase ev penetration in india,”
networks using hints,” in Proceedings of the Ninth International Confer- Energies, vol. 11, no. 3, p. 483, 2018.
ence on Machine Learning. Morgan Kaufmann, 1992, pp. 363–368. [112] D. Gupta and A. Garg, “Sustainable development and carbon neutrality:
[91] C. L. Giles, C. Omlin, and K. K. Thornber, “Equivalence in knowledge Integrated assessment of transport transitions in india,” Transportation
representation: Automata, recurrent neural networks, and dynamical Research Part D: Transport and Environment, vol. 85, p. 102474, 2020.
fuzzy systems,” Proceedings of the IEEE, vol. 87, no. 9, pp. 1623–1640, [113] G. K. Sarangi, “Green energy finance in india: Challenges and solu-
1999. tions,” ADBI Working Paper Series, Tech. Rep., 2018.
[92] S. Hochreiter, “The vanishing gradient problem during learning recur- [114] “The influence of increased population density in china on air pollution,”
rent neural nets and problem solutions,” Int. J. Uncertain. Fuzziness Science of The Total Environment, vol. 735, p. 139456, 2020.
Knowl.-Based Syst., vol. 6, no. 2, pp. 107–116, 1998. [115] Z. Zhang, T. Xue, and X. Jin, “Effects of meteorological conditions
[93] D. P. Kingma and J. Ba, “Adam: A method for stochastic optimiza- and air pollution on covid-19 transmission: Evidence from 219 chinese
tion,” in 3rd International Conference on Learning Representations, cities,” Science of The Total Environment, vol. 741, p. 140244, 2020.
ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track [116] R. Chandra, K. Jain, R. V. Deo, and S. Cripps, “Langevin-gradient paral-
Proceedings, 2015. lel tempering for Bayesian neural learning,” Neurocomputing, vol. 359,
[94] M. Schuster and K. Paliwal, “Bidirectional recurrent neural networks,” pp. 315–326, 2019.
Signal Processing, IEEE Transactions on, vol. 45, pp. 2673 – 2681, 12 [117] R. Chandra and A. Kapoor, “Bayesian neural multi-source transfer learn-
1997. ing,” Neurocomputing, vol. 378, pp. 54–64, 2020.
[95] A. Graves and J. Schmidhuber, “Framewise phoneme classification with [118] R. Chandra, Y.-S. Ong, and C.-K. Goh, “Co-evolutionary multi-task
bidirectional lstm and other neural network architectures,” Neural net- learning with predictive recurrence for multi-step chaotic time series

17
 prediction,” Neurocomputing, vol. 243, pp. 21–34, 2017.
[119] R. Chandra, A. Jain, and D. S. Chauhan, “Deep learning via LSTM
models for COVID-19 infection forecasting in india,” arXiv preprint
arXiv:2101.11881, 2021.

18