The Impact of COVID-19 Lockdown on Ambient Air Quality in Shanghai, 2022

Abstract

The COVID-19 lockdown contributes to the improvement of air quality. Most previous studies have attributed this to the reduction of human activity while ignoring the meteorological changes, this may lead to an overestimation or underestimation of the impact of COVID-19 lockdown measures on air pollution levels. To investigate this issue, we propose an XGBoost-based model to predict the concentrations of PM_2.5 and PM₁₀ during the COVID-19 lockdown period in 2022, Shanghai, and thus explore the limits of anthropogenic emission on air pollution levels by comprehensively employing the meteorological factors and the concentrations of other air pollutants. Results demonstrate that actual observations of PM_2.5 and PM₁₀ during the COVID-19 lockdown period were reduced by 60.81% and 43.12% compared with the predicted values (regarded as the period without the lockdown measures). In addition, by comparing with the time series prediction results without considering meteorological factors, the actual observations of PM_2.5 and PM₁₀ during the lockdown period were reduced by 50.20% and 19.06%, respectively, against the predicted values during the non-lockdown period. The analysis results indicate that ignoring meteorological factors will underestimate the positive impact of COVID-19 lockdown measures on air quality.

1. Introduction

The Yangtze River Delta, as one of the four source areas of heavy haze events in China, has led to considerable air pollution problems over the past few decades due to its rapid economic development and urbanization [1,2]. During the past three years from 2020 to 2022, some social distancing measures were adopted to control the COVID-19 epidemic in China, such as movement restrictions, the prohibition of public meetings, the shutdown of schools and public places, isolation, and quarantine [3]. These restrictions have significantly reduced anthropogenic emissions and prevented severe air pollution incidents in the Yangtze River Delta [4], including the reduced emissions from public and private vehicle transport [5]. Besides, some other major social events associated with the air pollution control measures, such as the Asia-Pacific Economic Cooperation (APEC) meeting in 2014 and the Victory Day parade of China in 2015, have demonstrated the positive effects of air pollution control measures [6]. A case study also indicated that Guangzhou’s ambient air quality was significantly improved in 2016 after the implementation of a series of air pollution control measures in Guangzhou and the surrounding cities from 2014 to 2016 [7].

Many countries have adopted closure measures during the COVID-19 epidemic period which control carbon emissions by sacrificing international shipping et al. and were mostly reported to have improved the local air quality [8,9,10]. As vehicle emissions and industrial productions are the main sources of PM_2.5 and PM₁₀ air pollution, these pollution sources were controlled by the safety measures adopted for the prevention of COVID-19. The impact of restriction policies for the COVID-19 pandemic on air quality varies considerably between countries depending on their different geographical, economic, industrial, and social characteristics [11]. By comparing the emission changes of the main pollutants in different regions of the world during the first six months of 2020, the eastern regions of China have the largest decrease in total emissions in February [4,12]. Research has shown that the concentrations of PM₁₀ decreased by 75% during the lockdown in Salé City (Morocco) compared to PM₁₀ in the normal period [13], and the same value decreased by 18% for PM₁₀ in Portugal [14]. As for the reduction of PM_2.5, the reports showed that the concentrations of PM_2.5 decreased 76.5% in Malaysia, 58% in Spain, 53.1% in Delhi, and 53.1% in Tehran [9]. In addition, a reduction of 29.8% in PM_2.5 was found during the partial lockdown period in São Paulo state, Brazil [15]. In the United States, PM_2.5 declined in urban counties with early closed businesses [16]. In Wuhan, China, the concentrations of PM_2.5 and PM₁₀ in the lockdown period declined by 41.2% and 33.1%, respectively, compared to the pre-COVID-19 period [17]. Besides, reductions in surface PM_2.5 and PM₁₀ were reported in six mega-cities in India during the lockdown period [18,19]. In 2020, PM concentrations at the traffic and background stations are on average 15% and 13.4% lower than in the reference period in Wrocław, Poland [20]. In contrast, in Italy, no decrease in particulate matter (PM₁₀ or PM_2.5) was observed during the first phase of the COVID-19 epidemic [21] and Few sites (2–3 out of 16) experienced statistically significant drops in PM_2.5 in Ontario, Canada [22].

However, most of the available studies on outbreak control measures and environmental pollutant levels have yet to quantify the effects of meteorological changes on air pollutant levels during the COVID-19 pandemic around the world, which would lead to an overestimation or underestimation of the impact of COVID-19 lockdown measures on air pollution levels. In general, meteorological factors are of special importance in affecting the regional air quality in China [1,23,24,25,26]. Meteorological conditions may be a determining factor in the occurrence of air pollution if the emissions of air pollutants in a region are essentially stable over a certain period [25]. Many meteorological factors, such as wind speed, wind force, temperature, precipitation, and humidity, also significantly affect air pollution levels [27,28,29]. The meteorological contribution to PM_2.5 levels ranged from −4.8% to 4.3% in the Yangtze River Delta region [30]. The PM₁₀ concentration is negatively correlated with the relative humidity and temperature [31]. The correlations between PM and NO₂ and between PM and SO₂ were moderate, while the correlations between PM and CO and between PM and O₃ were weak and unstable, respectively [32]. Unfavorable meteorological conditions can offset the positive environmental impacts of closure measures; one report stated that control measures reduced PM_2.5 in Beijing by 12 µg/m³. In comparison, meteorological forcing increased PM_2.5 by 30 µg/m³, resulting in an increase in PM_2.5 levels during the closure period [28]. In Hubei Province, China, which had the highest number of confirmed pneumonia cases in February 2020, the impact of total emissions reductions (72%) significantly outweighed the proportion of meteorologically driven PM_2.5 increases (13%) [21]. The average PM_2.5 concentration in the Singapore region decreased by 19% when the effect of meteorological parameters was taken into account; this decrease was lower than the decrease observed when the effect of meteorological parameters was not taken into account (29%) [33].

Although individual studies have incorporated meteorological parameters’ effects when assessing the impacts of COVID-19 on NO₂ and PM_2.5 pollutant levels [33], the machine-learning methods applied in those studies are inferior to XGBoost models in terms of their predictive accuracy [34,35]. XGBoost, as a typically integrated algorithm, can be executed efficiently using fewer resources [36]. Moreover, the XGBoost model combines hourly predictions with near-surface measurements of air pollutants and meteorological conditions to estimate daily PM_2.5 concentrations in Shanghai, China, with improved model correlation coefficients of 50–100% and reduced standard deviations of 14–24 µg/m³ compared to the WRF-Chem model [37]. In addition to using meteorological data, air quality standard data such as SO₂, NO_X, and O₃ data are utilized in the XGBoost algorithm to model the atmospheric PM₁₀ concentrations [38].

Environmental factors play an important role in the spread of coronavirus disease 2019 (COVID-19) [39]. A comprehensive crisis management strategy for pandemic threats must also be based on environmental and socioeconomic factors and new technologies, not just on medically relevant parameters [40]. To gain experience that can be used to prevent future pandemic threats and to avoid overestimating or underestimating the impact of the COVID-19 outbreak on air pollution, this study used meteorological parameters (the mean temperature, dew point, visibility, and wind speed) and the concentrations of the remaining air pollutants to estimate the level of PM in the air during the COVID-19 outbreak. In addition, to account for the unpredictability of the lockdown measurements and increase the comparability of the results, this study examined trends of air pollutant concentrations from 2020 to 2022 (i.e., following the COVID-19 epidemic), as opposed to limiting the analysis to a predetermined lockdown period. We also compare the changes in air pollution levels in Shanghai during the lockout period (31 March to 1 June) in each year between 2015 and 2022 to provide a more comprehensive understanding of the effects of the novel coronavirus on air pollution.

2. Materials and Methods

2.1. Sample and Data

Shanghai, China’s worldwide economic, financial, trading, shipping, science, and technology innovation center, is located between 120°52′ and 122°12′ east and between 30°40′ and 31°53′ north. In late February 2022, a mutated strain of the Omicron virus triggered a significant public health disaster in Shanghai. The Shanghai City Government initiated a series of containment and control measures on 31 March 2022, and by 1 June 2022, the epidemic situation in Shanghai had been effectively contained. This study takes the COVID-19 lockdown period in Shanghai, in 2022, as the case study, to investigate the impact of the COVID-19 lockdown on the air quality changes. The different periods employed in this study for the prevalence of COVID-19 in Shanghai are shown in

Table 1. The different periods employed in this study for the prevalence of COVID-19 in Shanghai.

Periods	Prevalence Control of COVID-19 in Shanghai
1 January 2015–31 December 2019	Period before COVID-19
1 January 2020–1 June 2022	Period during COVID-19
31 March 2022–1 June 2022	COVID-19 lockdown period

. The spatial distribution of Shanghai and monitoring stations is shown in Figure 1.

The daily Air Quality Index (AQI), which is used by government agencies to report daily air quality [41] and daily concentrations of six related air pollutants, PM_2.5, PM₁₀, NO₂, CO, SO₂, and O₃, in Shanghai from 1 January 2015 to 30 September 2022 were obtained from the national real-time urban air quality release platform of the China National Environmental Monitoring Centre at http://106.37.208.233:20035, accessed on 24 October 2022.

Table 2. Descriptive statistics of air quality data.

Parameter	Unit	Range	Average	St. Dev.
AQI		[10, 270]	61.96	32.19
PM2.5	μg/m3	[2, 219]	37.39	26.12
PM10	μg/m3	[6, 308]	52.07	31.47
NO2	μg/m3	[4, 142]	39.56	18.81
CO	mg/m3	[0.31, 2.21]	0.71	0.24
SO2	μg/m3	[3, 74]	9.31	6.20
O3	μg/m3	[8, 178]	70.57	28.73

shows the descriptive statistics of these air quality data. CO is given in mg/m³, the other gaseous pollutants are given in μg/m³ and AQI is a dimensionless relative value that collectively indicates the level of air pollution or the air quality rating.

The daily mean temperature (TEMP), dew point temperature (DEWP), visibility (VISIB), and wind speed (WDSP) data for Shanghai from 1 January 2015 to 30 September 2022 were obtained from the National Climatic Data Center (NCDC), which is part of the National Oceanic and Atmospheric Administration (NOAA) at Daily Observational Data (noaa.gov, accessed on 24 October 2022). The daily mean temperatures (TEMP) and dew point temperatures (DEWP) are given in degrees Fahrenheit; the daily mean visibility (VISIB) is given in feet; and the daily mean wind speed (WDSP) is given in knots.

2.2. Measures of Variables

2.2.1. Influence of Meteorological and Other Pollution Parameters on PM_2.5 and PM₁₀

The important factors that affect the migration and transformation of PM_2.5 and PM₁₀ include weather conditions and the concentrations of other pollutants [42]; moreover, numerous parameters influence the PM_2.5 and PM₁₀ concentrations and these processes are quite complex. Before estimating the PM_2.5 and PM₁₀ concentrations, it is crucial to study the essential parameters to improve the predictions and determine the change in these components’ potential to influence these concentrations. As not all distributions of the analyzed variables show normal distributions, the Spearman correlation coefficient [43] was chosen as the measuring indicator to evaluate the meteorological parameters and other pollution characteristics associated with PM_2.5 and PM₁₀ concentrations. The Spearman correlation coefficient was calculated using Equation (1), where x represents the rank variables of PM_2.5 or PM₁₀, y represents the rank variables of other data, and n represents the number of each type of data.

$${\rho }_{x,y}=\frac{n{{\displaystyle \sum }}_{i=1}^n{x}_i{y}_i-{{\displaystyle \sum }}_{i=1}^n{x}_i{{\displaystyle \sum }}_{i=1}^n{y}_i}{\sqrt{n{{\displaystyle \sum }}_{i=1}^n{x}_i^2-{\left({{\displaystyle \sum }}_{i=1}^n{x}_i\right)}^2}\sqrt{n{{\displaystyle \sum }}_{i=1}^n{y}_i^2-{\left({{\displaystyle \sum }}_{i=1}^n{y}_i\right)}^2}}.$$

(1)

2.2.2. XGBoost-Based Prediction Model

XGBoost (extreme gradient boosting) is a model that was initially proposed by Tianqi Chen and Carlos Guestrin in 2011 [36], which has been continuously optimized and enhanced by the follow-up researchers. The model is a learning framework built on boosting tree models. In view of this, we proposed an XGBoost-based prediction model to forecast the air quality changes during the COVID-19 lockdown period in Shanghai, 2022. The specific calculation steps of XGBoost-based prediction model are shown in Figure 2.

The model’s core concept is to combine the weak classifier CART tree with an additive model to create a strong classifier. Assuming that having a total of K trees, the tree model is integrated by using the addition method, with F representing the fundamental tree model:

$${\widehat{y}}_i={{\displaystyle \sum }}_{k=1}^K{f}_k\left(x_i\right),f_k\in F,$$

(2)

where ${\widehat{y}}_i$ represents the predicted PM concentration, $x_i$ represents the predictors, and $f_k\left(x\right)$ is the k-th CART (Classification and Regression Trees) loss function.

$$L={{\displaystyle \sum }}_{i}l\left({\widehat{y}}_i,y_i\right)+{{\displaystyle \sum }}_k\Omega \left(f_k\right),$$

(3)

where $l$ is the loss function, representing the error between the predictive value and the actual value; $\Omega$ is the function used for regularization to prevent overfitting. The term$\Omega$ is calculated as follows:

$$\Omega \left(f\right)=\gamma T+\frac{1}{2}\lambda \parallel w{\parallel }^2,$$

(4)

where $T$ represents the number of leaves per tree and $w$ represents the weight of the leaves of each tree.

This function can be simplified by using the Taylor expansion. A CART node formula for loss reduction following a tree split from a particular node can be derived as follows:

$$\mathrm{Gain} =\frac{1}{2}\left[\frac{{\left({{\displaystyle \sum }}_{i\in I_L}{g}_i\right)}^2}{{{\displaystyle \sum }}_{i\in I_L}{h}_i+\lambda }+\frac{{\left({{\displaystyle \sum }}_{i\in I_R}{g}_i\right)}^2}{{{\displaystyle \sum }}_{i\in I_R}{h}_i+\lambda }+\frac{{\left({{\displaystyle \sum }}_{i\in I}{g}_i\right)}^2}{{{\displaystyle \sum }}_{i\in I}{h}_i+\lambda }\right]-\gamma ,$$

(5)

where $I$ is a subset of the available observations in the current node, $I_L \mathrm{and} I_R$ are subsets of the available observations in the left and right nodes, respectively, following the split. The predictors are daily concentrations of four air pollutants (SO₂, NO₂, CO, and O₃) and daily meteorological data (temperature, dew point, visibility, and wind speed) observations. In this way, an XGBoost-based prediction model was proposed and used to predict the ambient PM concentrations from 2020 to 2022, which was regarded as the air quality without the impact of COVID-19 lockdown measures.

2.2.3. Accuracy Evaluation Measure

In this study, three indicators were chosen to evaluate the accuracy of the PM value predictions, that is, the mean absolute error (MAE), root mean square error (RMSE), and mean percentage error (MPE). The formulas of these three evaluation indicators are shown in Equations (6)–(8):

$$\mathrm{RMSE}=\sqrt{\frac{1}{n}{{\displaystyle \sum }}_{i=1}^n{\left({\widehat{y}}_i-y_i\right)}^2},$$

(6)

$$\mathrm{MAE} =\frac{1}{n}{{\displaystyle \sum }}_{i=1}^n\left|\left({\widehat{y}}_i-y_i\right)\right|,$$

(7)

$$\mathrm{MPE}=\frac{100\%}{n}{{\displaystyle \sum }}_{i=1}^n\frac{{\widehat{y}}_i-y_i}{y_i}$$

(8)

where $y_i$ and ${\widehat{y}}_{\mathrm{i}}$ represent the observed and predicted PM concentrations, respectively, and $n$ refers to the total logarithm of data in the test set.

2.3. Model Flowchart and Data Analysis Procedure

Figure 3 shows the proposed model flowchart and data analysis procedure, which include the following three steps:

Step 1. For air quality data and meteorological data collected in Shanghai from 1 January 2015 to 30 September 2022, during the data preprocessing phase, if air quality data were missing for a particular day, all data for that day were deleted; for missing individual visibility (VISIB) data, the average of all remaining valid visibility (VISIB) data was used to fill in the missing values. After that, all data were divided into two parts according to time: before the COVID-19 outbreak (Time 2015–2019) and during COVID-19 epidemic (Time 2020–2022).

Step 2. In order to investigate the impact of COVID-19 lockdown on air quality, the AQI data for a total of 63 days, from the 31 March to the 1 June (the time period of the lockdown in Shanghai), were independently analyzed for each of the years from 2015–2022. In addition, the difference between observed and predicted PM values during the lockdown in Shanghai were calculated to show the impact of COVID-19 lockdown on air quality.

Step 3. To avoid overestimating or underestimating the impact of the COVID-19 outbreak on air pollution, meteorological parameters (the mean temperature, dew point, visibility, and wind speed) and the concentrations of the remaining air pollutants (NO₂, CO, O_3, and SO₂) were chosen as predictors to estimate the level of PM in the air during the COVID-19 epidemic. The XGBoost-based prediction model was selected as the prediction model: the training set was the data before the COVID-19 outbreak (Time 2015–2019) and the test set was the data during the COVID-19 epidemic (Time 2020–2022). In order to investigate whether predictors without meteorological factors overestimate or underestimate the impact of the COVID-19 lockdown, the two prediction results using different predictors were compared.

3. Discussion and Results

3.1. Air Quality Investigation before and during COVID-19 in Shanghai

In order to investigate the air quality changes during the COVID-19 lockdown period in Shanghai from 31 March 2022 to 1 June 2022, we collected the AQI data with a total of 63 days from 31 March to 1 June from the year 2015 to 2022, the daily AQI tendency was presented for each year. To facilitate the data comparisons, we, respectively compared the AQI data for the period before COVID-19 and the period during COVID-19, as shown in Figure 4 and Figure 5. As is shown, three different AQI classification levels are employed to evaluate the daily air quality for each year in the same period, the specific classification levels are shown in

Table 3. Air Quality Index Scale and Color Legend.

Levels of Concern	Values of Index
Good	0~50
Moderate	51~100
Unhealthy for Sensitive Groups	101~150
Unhealthy	151~200

. It is obvious that the AQI in Shanghai are fluctuated during the study period of each year, and the AQI during the COVID-19 epidemic period (31 March to 1 June, 2020–2022) is notably lower than that during the period before COVID-19 (2015–2019).

To further compare the air quality quantitatively before and during COVID-19 in Shanghai, we statistically analyzed the numbers and percentages of days with different air quality levels in the lockdown period (31 March to 1 June) in Shanghai from 2015 to 2022. As it is presented in the

Table 4. Numbers and percentages of days with different air quality levels in the lockdown period (31 March to 1 June) in Shanghai from 2015 to 2022.

Year	Good/day	Proportion	Moderate/day	Proportion	Unhealthy for Sensitive Groups/day	Proportion	Unhealthy/day	Proportion
2022	45	71.43%	18	28.57%	0	0.00%	0	0.00%
2021	41	65.08%	18	28.57%	4	6.35%	0	0.00%
2020	26	41.27%	35	55.56%	2	3.17%	0	0.00%
2019	10	15.87%	48	76.19%	5	7.94%	0	0.00%
2018	10	15.87%	41	65.08%	12	19.05%	0	0.00%
2017	8	12.70%	48	76.19%	7	11.11%	0	0.00%
2016	8	12.70%	40	63.49%	12	19.05%	3	4.76%
2015	6	9.52%	49	77.78%	7	11.11%	1	1.59%

, the proportions of days with good air quality during the COVID-19 period in the year 2020, 2021, and 2022 are higher than 40%, which is significantly better than that (with the AQI lower than 16%) before the COVID-19 period (2015–2019). Furthermore, the number of days with good air quality during the COVID-19 period kept increasing every year, with 41.27%, 65.08%, and 71.43% in the years 2020, 2021, and 2022, respectively. It is noted that the epidemic lockdown period in 2022 are all with good or moderate air qualities in Shanghai. This can be attributed to the fact that the lockdown measures had a positive impact on the air quality since the COVID-19 broke out in December 2019.

3.2. Impacts of Prediction Factors on PM_2.5 and PM₁₀

To determine the prediction factors that influence the concentration of PM_2.5 and PM₁₀ during the COVID-19 epidemic period, we first classified the factors into two categories, that is, the meteorological parameters, including the temperature, dew point, visibility, wind speed, and other environmental pollutants, including the concentrations of SO₂, NO₂, CO, and O₃. To select the prediction factors with significant influence on the concentrations PM_2.5 and PM₁₀, the correlation coefficients between these factors and the concentrations of PM_2.5 and PM₁₀ are calculated and analyzed before and during the COVID-19 outbroken. The correlation coefficients are shown in

Table A1. Correlation coefficients between meteorological factors and other pollutant concentrations with PM_2.5 and PM₁₀ concentrations before and during the COVID-19 outbreak.

		SO2	NO2	CO	O3	Temperature	Dew Point	VISIB	Weed Speed
2015–2019	PM2.5	0.636 **	0.671 **	0.825 **	−0.033	−0.265 **	−0.308 **	−0.442 **	−0.282 **
	PM10	0.706 **	0.592 **	0.678 **	0.058 *	−0.202 **	−0.323 **	−0.216 **	−0.235 **
2020–2022	PM2.5	0.449 **	0.638 **	0.712 **	0.003	−0.297 **	−0.337 **	−0.373 **	−0.247 **
	PM10	0.603 **	0.557 **	0.461 **	0.107 **	−0.23 **	−0.376 **	−0.053	−0.143 **

* p < 0.05 ** p < 0.01.

and indicated with the confusion matrix in Figure 6 and Figure 7, respectively.

As is shown in Figure 6 and Figure 7, the selected prediction factors are significantly related with the concentrations of PM_2.5 and PM₁₀, with the p-values for the seven factors (temperature, dew point, visibility, wind speed, the concentrations of SO₂, NO₂, and CO) are lower than 0.01. It is noted that there exist substantial connections between O₃ and PM_2.5 according to Lin Huang’s research [44], although the correlation coefficient for O₃ was relatively low compared to those of SO₂, NO₂, and CO. Public researches indicated that high temperatures and high humidity would lead to the dust to be more active to disperse into the air, and, therefore, increase the concentrations of PM_2.5 and PM₁₀ [45]. Furthermore, weak wind is not conducive to the diffusion of air pollutants and usually results in higher concentrations of PM_2.5 and PM₁₀ [46]. Therefore, the combination of all these factors will comprehensively influence the concentrations of PM_2.5 and PM₁₀. In addition, our study also demonstrated that the correlation coefficients between the concentrations of PM_2.5 and PM₁₀ and other air pollutants and various meteorological factors did not change notably during the two different time periods (before the COVID-19 outbroken and during the COVID-19 epidemic), as compared by Figure 6 and Figure 7. Therefore, all factors were selected to predict the concentrations of PM_2.5 and PM₁₀ during the COVID-19 epidemic period.

3.3. Analysis of the Predicted PM_2.5 and PM₁₀ Concentrations in Shanghai

3.3.1. Data Prediction and Accuracy Analysis

The XGBoost-based prediction model was used to predict concentrations of PM_2.5 and PM₁₀ in Shanghai. As for the model with the dependent variable PM_2.5, the predictors are the daily concentration data of NO₂, CO, SO₂, and O₃ and daily mean temperature, dew point temperature, visibility, and wind speed data. In terms of the model with the dependent variable is PM₁₀, the predictors are the same as that of PM_2.5 prediction. Additionally, the data before the COVID-19 outbreak (2015 to 2019) was set as the training set, which was used to predict the daily PM_2.5 and PM₁₀ concentrations during the period of the COVID-19 epidemic (2020 to 2022), and the results of the predicted and observed values were obtained as shown in Figure 8 and Figure 9. Furthermore, the root mean square error (RMSE) and mean absolute error (MAE) of the PM_2.5 and PM₁₀ predictions for each year were calculated and are shown in

Table 5. Evaluation of the prediction results for each year during the COVID-19 epidemic.

	Year	RMSE	MAE
PM2.5	2022	12.104	8.590
	2021	11.843	8.419
	2020	11.705	8.691
PM10	2022	17.390	12.545
	2021	21.423	14.249
	2020	17.448	11.935

.

As is shown in Table 5, the root mean square error (RMSE) of PM_2.5 predictions for each year are all distributed around 12, which is lower than other PM_2.5 prediction accuracy results from existing studies using the same model [39,47,48]. Moreover, the root mean square error (RMSE) and the mean absolute error (MAE) of PM₁₀ predictions for each year are all under 21.4 and 14.3, respectively, which are lower than other PM₁₀ prediction accuracy results using different predicted models like MLR and ANN [49].

3.3.2. Comparison of Predicted and Observed Value during the Lockdown Period in Shanghai

To investigate the effect of COVID-19 lockdown measures in Shanghai on the PM_2.5 and PM₁₀ concentrations, statistical comparisons of the differences between predicted and observed concentrations of PM_2.5 and PM₁₀ were performed from 31 March to 1 June 2022. The results are presented in

Table 6. Statistical comparisons of the differences between predicted and observed concentrations during the lockdown period, with * indicates that p-value is lower than 0.05.

	Z	p
PM2.5	−3.821	0.000 *
PM10	−3.000	0.003 *

. Furthermore, the difference between the observed and predicted concentrations of PM_2.5 and PM₁₀ during the lockdown period are shown in Figure 10 and Figure 11.

As is presented in Table 6, the predicted and observed values of PM_2.5 and PM₁₀ differ significantly during the lockdown period, as the p-value were 0.000 and 0.003. From this, we can speculate that lockdown measures have a significant impact on the daily concentration of PM_2.5 and PM₁₀. In addition, as is shown in Figure 10 and Figure 11, the majority of predicted PM_2.5 and PM₁₀ values were higher than observed values during the lockdown period. It is calculated that the mean value of the subtractive difference between the observed and predicted PM_2.5 concentrations is −8.49 μg/m³, and this value is −7.51 μg/m³ for PM₁₀. Moreover, the predicted concentrations of PM_2.5 and PM₁₀ in Shanghai were higher by an average of 60.81% and 43.12% per day, respectively compared to the observed values throughout the lockdown period from 31 March to 1 June 2022. Therefore, COVID-19 lockdown measures in Shanghai can be considered to have a positive impact on air quality, which is similar to the findings of some public studies researching the changes in the levels of PM_2.5 and PM₁₀ during the lockdown [9,15,17,18,19,20]. However, unlike other studies, our study regarded predicted concentrations of PM_2.5 and PM₁₀ during the lockdown period as the concentrations without lockdown measures.

3.3.3. Comparison of Prediction Results with and without Meteorological Predictors

To investigate whether predictors without meteorological factors overestimate or underestimate the impact of the COVID-19 lockdown, Holt-Winters time series exponential smoothing model [50] was used to predict the concentrations of PM_2.5 and PM₁₀ during the lockdown period. In the model where the dependent variable is PM_2.5, the predictors are the concentrations of PM_2.5 before COVID-19 broke out. In the model where the dependent variable is PM₁₀, the predictors are the concentrations of PM₁₀ before COVID-19 broke out. The results of PM_2.5 and PM₁₀ predicted values by using different prediction models are shown in Figure 12 and Figure 13. Furthermore, the mean percentage error (MPE) of PM_2.5 and PM₁₀ predictions during the lockdown period in Shanghai using XGBoost-based prediction model and Holt-Winters time series prediction model are shown in

Table 7. The mean percentage error (MPE) of PM_2.5 and PM₁₀ predictions during the lockdown period in Shanghai by two different prediction models.

	XGBoost-Based Prediction Model	Holt-Winters Time Series Prediction Model
PM2.5	60.81%	50.20%
PM10	43.12%	19.06%

.

As is indicated in Figure 12 and Figure 13, the prediction results by using meteorological factors as predictors (XGBoost-based prediction model) have a better presentation of the volatility of daily PM_2.5 and PM₁₀ concentrations. It can be noted that most predictions without meteorological predictors are lower than those with meteorological predictors. In addition, as is shown in Table 7, it is calculated that the predicted concentrations of PM_2.5 and PM₁₀ without meteorological predictors (Holt-Winters time series prediction model) were higher by an average of 50.20% and 19.06% per day, respectively compared to the observed values throughout the lockdown period from 31 March to 1 June 2022 in Shanghai. In contrast, the predicted concentrations of PM_2.5 and PM₁₀ using meteorological predictors (XGBoost-based prediction model) were 60.81% and 43.12% higher on average per day compared to the observed values. Therefore, if the pre-COVID-19 epidemic data were used to predict PM_2.5 and PM₁₀ concentrations during the COVID-19 lockdown period without the use of meteorological predictors, the positive impact of the lockdown measures on air quality would be underestimated.

4. Conclusions

This study investigated the ambient air quality changes in Shanghai during the COVID-19 lockdown period. Firstly, by analyzing the air quality during the study period, it was found that the proportion of days with good air quality in the three years of COVID-19 were increased year by year, with 41.27%, 65.08%, and 71.43% in 2020, 2021, and 2022, respectively. Then, based on the proposed XGBoost-based time series prediction model, the concentrations of PM_2.5 and PM₁₀ are predicted with the assumption that no COVID-19 lockdown measures were adopted. Thus, by comparing the actual observations during the COVID-19 lockdown period with predicted values, the concentrations of PM_2.5 and PM₁₀ were totally reduced by 60.81% and 43.12% in COVID-19 lockdown period (from 31 March to 1 June 2022). Furthermore, in terms of the predicted results without considering meteorological factors, the concentrations of PM_2.5 and PM₁₀ during the lockdown period were reduced by 50.20% and 19.06%, respectively, compared to the predicted values during the non-lockdown period. These indicate that the epidemic containment measures have had a positive impact on air quality.

To sum up, this study mainly provides three contributions: (1) quantified the continuous air quality changes and compared air pollution levels before and during the COVID-19 lockdown period in Shanghai; (2) comprehensively utilized the meteorological factors and air pollutant data to predict the concentration of PM in Shanghai during COVID-19 lockdown period in 2022, which has avoided the underestimation of the positive impact of COVID-19 lockdown measures on air quality by considering the meteorological factors; (3) Under the most basic conditions of life, the impact of human activity restrictions on PM provides data support for the future government to formulate environmental governance measures.

However, this paper is limited in some respects. In our future studies, we will try to use larger amounts of PM_2.5 and PM₁₀ data in more regions for better model performance. In this work, we only studied the changes in air quality before and during the COVID-19 lockdown period, except for this, our future studies will focus on the air quality changes before, during, and after the lockdown period. Furthermore, we did not consider the variations of PM_2.5 or PM₁₀ with specific human factors such as mobility, industry, manufacturing, or residential activity due to the lack of data. Therefore, our future studies will try to investigate correlations between the variations of PM_2.5, PM₁₀ and the specific anthropogenic factors.