Impact of COVID-19 lockdown on air quality analyzed through machine learning techniques

Data description

Data on atmospheric pollution for Lahore, Pakistan, was gathered from IQAir (https://www.iqair.com/us/pakistan/punjab/lahore). A firm in Switzerland that specializes in preserving and monitoring airborne contaminants is called IQAir. It also deals with establishing and overseeing air cleansing and quality. IQAir has received recognition for its air purifiers (Asarch & Glendon, 2022; PTPA Media, 2020b; International Housewares Association, 2019; Morgan, 2019; PTPA Media, 2020a) as well as for its finest air quality monitor (BK-E Magazine, 2022). Data from the biggest network of ground-based sensors is gathered by IQAir. To guarantee data accuracy and dependability, the organization has deployed automated and AI-based tools and sensors (air-quality meter, IOT-based air-quality monitoring system, MQ135 sensors, etc.). Phase8-DHA, Zeenat Block, Dawn Food Head Office, and HAC Agri Limited are the four stations from which data is gathered.

The hourly concentrations of various pollutants, such as PM2.5, PM10 (particulate matter), NO2 (nitrogen dioxide), NH3 (ammonia), and O3 are used to calculate the air quality data (ozone). Along with this, a few additional factors were gathered, including air temperature, air pressure, humidity, wind speed, and wind direction.

Data pre-processing

An hourly data on a variety of air contaminants was gathered from January 2018 to May 2020. The missing values were substituted by the mean of that column for data variables with missing values of 1,000 or less. January 2018 to May 2020 is covered by the total processed data. Table 1 lists station details as well as a list of the atmospheric contaminants and the number of missing values for each one. For the purpose of identifying the outlier and imputing its values to the column mean, we employed scatter plots and the Mahalanobis Distance (MD) to discover the outlier.

Machine learning techniques

In machine learning, we work with a variety of algorithms that enable us to discover the link between data to produce the final prediction; the type of prediction model where we need to find prediction output in the form of continuous numerical values is known as a regression issue. Modern machine learning techniques, such as CNN, N-BEATS, decision tree regressor, support vector regressor, random forest regressor, auto-regressive integrated moving-average (ARIMA), and LSTM, have been used to explore and analyze the changes in atmospheric pollutants that occurred during the COVID-19 lockdown.

Mathematical model

Let LSTM has input x(t) and the output o(t)

h(t-1) and c(t-1) these are inputs from previous time step

c(t) and h(t) are consumption of next time-step, as shown in Fig. 1.

(1)${\displaystyle f_t={\sigma }_{\mathit{g}}\left({\mathit{W}}_{\mathit{f}}\times {\mathit{x}}_{\mathit{t}}+{\mathit{U}}_{\mathit{f}}\times {\mathit{h}}_{\mathit{t}-1}+{\mathit{b}}_{\mathit{f}}\right)}$ (2)${\displaystyle i_t={\sigma }_{\mathit{g}}\left({\mathit{W}}_{\mathit{i}}\times {\mathit{x}}_{\mathit{t}}+{\mathit{U}}_{\mathit{i}}\times {\mathit{h}}_{\mathit{t}-1}+{\mathit{b}}_{\mathit{i}}\right)}$ (3)${\displaystyle o_t={\sigma }_{\mathit{g}}\left({\mathit{W}}_{\mathit{o}}\times {\mathit{x}}_{\mathit{t}}+{\mathit{U}}_{\mathit{o}}\times {\mathit{h}}_{\mathit{t}-1}+{\mathit{b}}_{\mathit{o}}\right)}$ (4)${\displaystyle c{\stackrel{ˆ}{}}_t={\sigma }_{\mathit{c}}\left({\mathit{W}}_{\mathit{c}}\times {\mathit{x}}_{\mathit{t}}+{\mathit{U}}_{\mathit{c}}\times {\mathit{h}}_{\mathit{t}-1}+{\mathit{b}}_{\mathit{c}}\right)}$ (5)${\displaystyle c_t={\mathit{f}}_{\mathit{t}}.{\mathit{c}}_{\mathit{t}-1}+{\mathit{i}}_{\mathit{t}}.\left(c{\stackrel{ˆ}{}}_t\right)}$ (6)${\displaystyle h_t={\mathit{o}}_{\mathit{t}}.{\sigma }_{\mathit{c}}\left({\mathit{c}}_{\mathit{t}}\right)}$ Where

f_t is the forget gate

i_t is the input gate

o_t is the output gate

c_t is the cell state

h_t is the hidden state

σ_g is the sigmoid

σ_c is the tanh

. is the element wise multiplication

LSTM equations also generated f(t), i(t) and c $\stackrel{ˆ}{}$ these are the internal consumption of LSTM and being used for generating c(t) and h(t).

The above mathematical model shows the identical steps for one time-step, meaning all the equations must be recomputed for next time-step. Suppose we have 10 time-steps, we need to recompute the above equation for ten times. (Wf, Wi, Wo, Wc, Uf, Ui, Uo, Uc) all these are weight matrices and (bf, bi, bo, bc) are the biases and these are not time dependent.

Exploratory analysis of data

The LSTM model was trained using data from January 2018 to December 2019. But a different collection of special data from January to February 2020, known as the validation set, was utilized for validation (VS). The lock-down set was the data from March 2020 to May 2020. (LD). There are 3,647 instances in the validation and lock-down collection, which represent the hourly AQI values of various air contaminants. To estimate the values for VS and LD, the already trained model was employed.

In order to forecast the concentration of particular contaminants, the LSTM model was used. Target and predictive variables serve as the foundation for machine learning and other deep learning approaches  (Šimić et al., 2020). The concentration of certain air pollutants such PM2.5, PM10 (particulate matter), NO2 (nitrogen dioxide), and O3 were the goal variables (y) (ozone). These predictive factors helped the LSTM model capture the seasonal trend of traffic flow and industrial operations. The predictive variables (X) were data linked to environmental conditions coupled with their temporal variables. The RMSE, MAE, and R-SQUARE values were used to assess the model’s performance.

The classification of datasets into training, validation, and lockdown time frames is shown in Fig. 2 along with dataset details. The LSTM model was trained using time series data from January 2018 to December 2019, the validation set was used for cross-validation from January 2020 to February 28, 2020, and the actual values during the lockdown period (March–May) were compared with the model predictions using least square errors. The entire research methodology is shown in Fig. 3. First, the dataset is preprocessed, then thoroughly refined, and preprocessed data is passed to the most cutting-edge machine learning models. Next, each of the machine learning models is evaluated based on RMSE, MAE, and R-SQUARE score, and finally the final results are compared. The model with the lowest error score is considered to be the best model.

Inferential statistical analysis

To infer attributes from the dataset, inferential statistical analysis is employed. These characteristics aid in predicting future values. Different hypotheses can be tested; inferential statistics heavily relies on the testing of hypotheses. It enables us to judge if the pattern we notice is the result of chance or whether it actually has statistical significance. In order to achieve this, we must first create a hypothesis and a null hypothesis. The hypothesis is the trend we expect to see in the data, while the null hypothesis is the polar opposite. In our investigation, we focused on four air pollutants: PM2.5, PM10, NO2, and O3. We will analyze how these contaminants impact the air quality index (AQI). According to our theory, the concentration of these contaminants and the AQI are linearly correlated. The null hypothesis demonstrates that there is no linear relationship between these pollutants’ concentrations and AQI. Following that, we do various tests to determine if there is statistical significance. Table 2 demonstrates that the p-value for our parameters is less than 0.05 and that the F-statistics for PM2.5 is substantial. The null hypothesis may thus be rejected as we can presume that there is statistical significance. Similar circumstances apply to the pollutants PM10, NO2, and O3. In Table 2, the negative slope indicates that the air quality index (AQI) will decline as pollutant concentration rises and increase as concentration falls.

Validation method and comparison

The projected concentration of each air contaminant is shown by the LSTM findings. These predicted concentrations and the contaminants’ actual values were contrasted. Both from a long-term and short-term viewpoint, the evolution of the historical data was compared to our technique. Due to a decrease in traffic and industrial activity, we see a decrease in the concentration of air pollutants.

Explorative analysis

We use explorative data analysis to determine the relative importance of each air contaminant in order to comprehend the relationships between them. We employ principal component analysis for that (PCA). For the best outcomes, it is advised to select the components that account for between 70 and 80 percent of the overall variation based on explained variance. We select five elements from the data that are the most beneficial and also include important information by keeping the concept in mind.

Hyper-parameter tuning for LSTM

To influence the behavior of deep learning and machine learning models, hyper-parameter tweaking is crucial. Our model’s ability to produce accurate results depends on our ability to fine-tune its hyper-parameters. If we are unable to minimize the loss-function, our model will produce more mistakes. We have adjusted the hyper-parameters for the LSTM model while keeping this idea in mind. The LSTM model has four layers: one input layer, two hidden layers, and one output layer. It is generally recommended to utilize two hidden layers for complicated problems, and we have done the same. Second, we concentrate on the number of nodes. When we lower the number of nodes, it improves accuracy and, in other words, reduces overfitting. Removing nodes act as a regularizer. A total of 50 nodes were employed in the input layer, while 50% fewer nodes were used in the second and third layers. We used the ReLU activation function on the hidden layers because ReLU is nonlinear and has the advantage of not having back propagation error, and we used the tanh activation function on the output layer because the output from that layer ranges from −1 to 1, and tanh is thought to be a good one for solving regression problems. In our situation, after 100 epochs the validation accuracy started to fall, therefore, we set epochs to 100. As Adam optimizer incorporates the best qualities of RMSProp and AdaGrad algorithms, it can handle the sparse gradients on noisy data extremely quickly. The learning rate was set to 0.1. We utilized root mean square error and the Adam optimizer as measurements. Although we have used LSTM with a variety of settings, the results are greatest when using the fine-tuned parameters stated above.

Evaluation parameters

We employed several machine learning assessment criteria. RMSE, MAE, and R-SQUARE score were among the metrics used.

R-SQUARE score

The third parameter is the R-SQUARE score, defined as the statistical measures that represent the proportion of the variance for a dependent variable that is explained by an independent variable or a variable in a regression model. In contrast, the correlation shows a strong relationship between the dependent and independent variables. It also shows the extent to which a variable explains the variance of another variable. (9)${\displaystyle {\mathit{R}}^2={\mathit{SS}}_{\mathit{regression}}/{\mathit{SS}}_{\mathit{total}}}$ Where

SS_regression is the sum squares by the regression

SS_total is the total sum of squares

Deep learning results

We have trained various machine learning models, including decision tree, SVR, random forest, ARMIA, CNN, N-BEATS, and LSTM, on the entire training dataset, which contains historical records from January 2018 to December 2019, in order to better understand the true changes in the concentration of various atmospheric pollutants. Additionally, these models were fitted independently to the data from January 2020 to May 2020 (validation set and lockdown set). The results were then assessed using the mean absolute error (MAE), R-SQUARE Score, and root mean square error (RMSE). When compared to other models, the LSTM model performs best and has the fewest mistakes. The primary explanation for this is because LSTM excels in time-series forecasting and develops forecasts based on prior sequential data. Compared to other models, LSTM has a remarkable capacity to efficiently remember data.

Figure 4 displays the LSTM model’s training and validation losses in terms of mean square error and root mean square error. On one hundred epochs, the initial training was run. After 10 epochs, the training and validation losses are consistent. During the first stages of training, the error was at its highest point and the gap between the two losses was at its greatest.

Comparison of LSTM model with other machine learning models

In relation to various pollutants, the overall findings of several machine learning models are shown in Table 3. On the air quality dataset, we applied a variety of machine learning models, including decision tree, SVR, random forest, ARIMA, CNN, N-BEATS, and LSTM. Error metrics including RMSE, MAE, and R-SQUARE score were used to evaluate the machine learning models. The top four pollutants accounted for 70% of the overall variance when we used PCA to choose features. The contaminants that were chosen were PM2.5, PM10, NO2, and O3. For each pollutant, we have utilized a scatter plot and the Mahalanobis Distance (MD) to identify the outlier and impute its values using the meaning of the column. For each pollutant, we modified the LSTM model’s parameters and put other cutting-edge machine learning models into practice. The findings were then created based on error measurements. All of the machine learning models worked well, however the LSTM model excelled in comparison to the others. The LSTM model has the lowest error for each of the various pollutants when compared to other models. Since we are projecting the pollutant concentration and utilizing time series data, the LSTM network is the ideal model for time series data. LSTM has the capacity to predict future values based on prior sequential data (past historical records). When compared to other machine learning models, LSTM performed better since it can anticipate the future while learning the prior levels of various contaminants. PM2.5 pollutant had a score of 4.46 percent MAE, 0.25 percent R-SQUARE score, and 9.99 percent RMSE from the LSTM model. The PM10 pollutant’s LSTM score was 12.84 percent RMSE, 0.75 percent R-SQUARE score, and 8.58 percent MAE. The results of the LSTM model for NO2 were 4.35 MAE, 0.12 R-SQUARE, and 6.77 RMSE. 8.20 percent MAE, 0.17 percent R-SQUARE, and 29.79 percent RMSE made up the O3 LSTM score.

The four parts that make up an elementary LSTM unit are as follows: a cell, an input gate, an output gate, and a forget gate. A cell’s function is to retain values for any given period of time. Additionally, gates’ function is to control how information enters and exits a cell. For time-series forecasting, LSTM consistently delivered good results. Sequential data’s future values are effectively predicted. There is no need for precise modification because the LSTM gives us a wealth of parameters, including input, output, biases, and learning rates. The LSTM model outperforms ARMIA in terms of performance. The major causes of this are because ARMIA is computationally costly and performs poorly for situations involving seasonal time-series and long-term forecasting. While LSTM often works with large datasets, where sufficient data must be provided for initial model training, ARIMA excels on smaller datasets, but it might be challenging to pinpoint the ARIMA model’s pivotal moments. When it comes to decision trees and random forests, a forest needs a lot of memory since it stores data from hundreds of trees. The LSTM cell stores information, and gates control that memory. Similarly, visualize the series of choices that results in computational burden. That is a reliable answer. The SVM method works well for small datasets, but it struggles with noisy data, data where target classes overlap, or data where the number of training data samples is insufficient for the number of features. As a result, these algorithms do not perform as well as LSTM for air quality datasets.

LSTM results with respect to different pollutants

Figure 5 shows the PM2.5 concentration for the period from January to May. The predicted values, which were compared to the actual values. The time series data from January to May are displayed on the plot. The validation set is from January 2020 to February 28th, and the lockdown period is from March 2020 to May 2020. The graph’s analysis reveals that there is no average decrease in PM2.5 concentration from January to the end of February. However, there is a decrease in the concentration of PM2.5 after the middle of March (the lockdown timeframe).

The concentration of NO2 for the period from January to May is shown in Fig. 6. The predicted values, assessed against the actual values. The validation set is from January through February 28th, and the lockdown period is from March through May of the following year. The concentration of NO2 decreases from the middle of March to the end of May. The graph reveals a steady decline in air pollution after March, which appears to be the true result of the COVID-19 lockout.

Figure 7 displays the PM10 concentration. The predicted values, which were compared to the actual values. The figure displays the time series data as well as the PM10 concentration. The validation set is the period from January 2020 to February 2020; the lockdown period is from March 2020 to May 2020. It was noticed that the amount of PM10 pollutants was falling.

Figure 8 displays the ozone (O3) levels for various time periods. The predicted values, which were compared to the actual values. The statistics for several time periods from January through May are represented on the figure. The validation set is from January through February 28th, and the lockdown period is from March through May of the following year. The graphic unequivocally demonstrates that the amount of ozone is rising (O3). The quality of ozone has significantly improved as a result of the concentration of some air contaminants being lower. There is a decrease in the concentration of air contaminants, as seen in Figs. 3, 6 and 7.

Forecasting moving average

The first stage in the forecasting process is to ascertain the maximum number of hours the model can anticipate accurately. Most of the time, the extended lead-time may cause performance that is ineffective and inaccurate. The R-SQUARE error levels determine how accurate the results are. R-SQUARE error levels increase as predicting time increases. On the other side, a shorter predicting period results in the least number of mistakes. In light of this, we have predicted the concentration of several pollutants, including PM2.5, PM10, NO2, and O3, over the next four days using a moving average.

The 4-day moving average of PM2.5 and PM10 pollutants is shown in Figs. 9A and 9B, and the findings show a reduction in PM2.5 and PM10 pollutant concentrations for the next four days. The error measurements have not been impacted by the moving average. The LSTM model still shows minimum errors in terms of RMSE, MAE, and R-SQUARE score while forecasting 4 days of moving average.

Figure 10A depicts a drop in NO2 concentration and Fig. 10B depicts an increase in O3 concentration (ozone). In terms of RMSE, MAE, and R-SQUARE score, the 4-day prediction of moving average produces strong results with little mistakes.

Reduction in air pollution

Comparing the lockdown of 2020 to the same time period in the preceding years, Table 4 provides a clearer picture of certain contaminants and their average concentrations.

We calculated the average concentration for each pollutant and displayed it in Table 4. For our research investigation, the well-known contaminants that were thought to be the more serious causes of pollution were selected. The contaminants that were considered were PM10, PM2.5, NO2, NOx, and O3. These contaminants’ typical concentrations were measured and contrasted with values from the same period in the prior years. Table 4 displays time series data for three months and compares the average concentration of various pollutants during the lockdown periods (March 2020 to May 2020) to the corresponding periods for the two years prior, 2018 and 2019, as well as the average concentration of various pollutants chosen for the years of 2018, 2019, and 2020. (March, May, and April). We found that the concentration of PM2.5 pollutants was 31% lower in March 2020 than it was in March 2019 after examining these statistical data.

When the PM2.5 pollutant concentrations in April (2019 and 2020) were compared, April 2020 had a 33 percent lower pollutant concentration than April 2019. When May 2019 and May 2020’s PM2.5 concentrations were compared, it was found that the latter month’s reading was 57 percent lower. Similar to March 2019, there was a 7.8% decrease in PM10 concentration in March 2020 compared to March 2019. When compared to April 2019, there was a 42 percent decrease in April comparing March 2020 to March 2019, the concentration of NO2 declined to 1.92 percent, by 43 percent in April 2020, and by 67 percent in May 2020. In comparison to March 2019, April 2020, and May 2019, the average NOx concentration reduced by 11.1 percent, 40 percent, and 70 percent, respectively. The quality of the air will improve as the number of atmospheric pollutants decreases, improving the overall concentration of ozone O3 during the lockdown period. In comparison to March 2019, there was a 38 percent rise in ozone (O3) in March 2020. Ozone (O3) levels increased for the month of April 2020 by 9.3% compared to April 2019. In addition, ozone (O3) increased by 12 percent in May 2020 compared to May 2019 within the same calendar year. When we compared the levels of several contaminants in 2020 and 2018, there was a considerable drop. Figure 11 provides a clear visual illustration of Table 4’s graphical form.

Figure 11 shows a visual depiction of the average concentration of several air pollutants during the months of March, April, and May for the years (2018), 2019 and 2020.