Airport terminal passenger forecast under the impact of COVID-19 outbreaks: A case study from China

Abstract

Passengers significantly affect airport terminal energy consumption and indoor environmental quality. Accurate passenger forecasting provides important insights for airport terminals to optimize their operation and management. However, the COVID-19 pandemic has greatly increased the uncertainty in airport passenger since 2020. There are insufficient studies to investigate which pandemic-related variables should be considered in forecasting airport passenger trends under the impact of COVID-19 outbreaks. In this study, the interrelationship between COVID-19 pandemic trends and passenger traffic at a major airport terminal in China was analyzed on a day-by-day basis. During COVID-19 outbreaks, three stages of passenger change were identified and characterized, i.e., the decline stage, the stabilization stage, and the recovery stage. A typical “sudden drop and slow recovery” pattern of passenger traffic was identified. A LightGBM model including pandemic variables was developed to forecast short-term daily passenger traffic at the airport terminal. The SHapley Additive exPlanations (SHAP) values was used to quantify the contribution of input pandemic variables. Results indicated the inclusion of pandemic variables reduced the model error by 27.7% compared to a baseline model. The cumulative numbers of COVID-19 cases in previous weeks were found to be stronger predictors of future passenger traffic than daily COVID-19 cases in the most recent week. In addition, the impact of pandemic control policies and passengers' travel behavior was discussed. Our empirical findings provide important implications for airport terminal operations in response to the on-going COVID-19 pandemic.

Graphical abstract

1. Introduction

The 2019 coronavirus disease (COVID-19) pandemic has unprecedentedly alerted people's life around the world since the first outbreak in early 2020 [1]. Working and studying from home has become the primary choice for many people amid the pandemic, which has influenced the operation and management of most non-residential buildings [[2], [3], [4]]. The use of transportation hub, such as airport terminals, has also been significantly alerted due to the reduced travel demands and the travel restrictions imposed by the government [5,6]. According to the International Civil Aviation Organization (ICAO), the total world air passenger traffic reduced by 60% in 2020 and by 49% in 2021 compared to 2019 [7]. The tremendous passenger uncertainty caused by the COVID-19 pandemic has created significant challenges for airport terminal operation and management. Developing optimized strategies for airport terminals to cope with pandemic-induced traffic fluctuations has been a topic of interest in recent studies [6,[8], [9], [10]].

Airport terminals are reported to consume approximately two to three times as much energy as typical non-residential buildings due to their functional and operational characteristics [11,12]. Improving energy efficiency while providing good indoor environmental quality (IEQ) is a primary objective for airport terminal operations and management [13,14]. Generally, more than 40% of the energy in airport terminals is consumed by the heating, ventilation, and air conditioning (HVAC) system to provide comfortable thermal environment for passengers [[15], [16], [17]]. The number of passengers is directly related to the demand for fresh air and the load of the air conditioning system [8,18]. Besides, operations of other airport service and applications such as lighting, elevator, hot water, baggage delivery, business service and ground handling service are also significantly affected by passenger patterns [6,19]. Energy consumption per passenger traffic has been used as an important indicator of airport energy efficiency in existing standards and studies [20]. Accurate passenger forecasting provides an essential foundation for optimizing the operation of HVAC and other building systems within an airport terminal building.

The pandemic has posed a huge challenge for future passenger forecasting. Traditionally, passenger profiles from previous years was considered an important baseline for forecasting [21]. In many existing studies, historical passenger traffic profiles have been used as the critical predictor of time series traffic forecasts in regression models and machine learning models [[22], [23], [24]]. Additional socioeconomic factors, such as national and regional gross domestic product (GDP), income levels, population, employment rates, and special events, were used to fine-tune the forecast [21,25]. Some studies have also attempted to predict travel demands with unconventional factors such as search engine queries [26,27] and social media data [28]. However, the outbreak of the COVID-19 pandemic has substantially increased the uncertainty of air travel demand, an effect that is difficult to address with the traditional predictors described above. It is necessary to include pandemic-related variables, such as the number of new cases, number of deaths, and pandemic control policies, in the forecasting model to account for the impact of the pandemic on air passenger [[29], [30], [31]]. However, few studies have been done on the above topics. More empirical studies are needed to explore which pandemic-related variables have strong predictive power for airport passenger traffic forecasts and the extent to which the inclusion of pandemic variables affects the accuracy of the model outcome.

Several studies have been undertaken to investigate air passenger traffic patterns in China during the COVID-19 pandemic to provide guidance to airports in dealing with these enormous challenges [[32], [33], [34], [35]]. However, most current studies focused only on the first outbreak of the pandemic during 2020 Spring, while the impact of later outbreaks has been largely overlooked. In the second half of 2020 and throughout 2021, outbreaks reoccurred intermittently in China, but with fewer cases and shorter durations during each subsequent outbreak, as shown in Fig. 1 . China reported a total of 35,107 COVID-19 cases in 2021, only 36% of the number in 2020, and none of the outbreaks were of similar magnitude to the first Wuhan outbreak in early 2020. Nevertheless, these intermittent outbreaks continued the depression in travel demand in 2021, with total air passenger traffic increasing by only 5.5% over 2020 [36]. Given the updates in pandemic control measures, evaluations of the virus and changes in people's psychological responses to the virus, little is known about whether these small-scale outbreaks yielded impacts on passenger traffic similar to those of the first major outbreak.

Fig. 1.

New daily confirmed cases of COVID-19 in China (data source: World Health Organization database, https://covid19.who.int).

The acquisition of data is another limitation frequently mentioned in existing studies. Several open access and subscription databases were cited as the primary data sources in relevant studies, such as the Civil Aviation Administration of China (CAAC) database, the International Air Transport Association (IATA) database and the Official Airline Guide (OAG) database [33,37,38]. However, data on passenger traffic in China are not often available. The time series of the passenger data are often incomplete or have a low resolution, i.e., monthly data, which imposes many limitations on the analysis. As an alternative, the number of flights and available seats were most often used as indicators to characterize travel demands in the current study. Even so, the number of passengers is a more accurate indicator of travel demand than either of these two.

To address the limitations of current studies, this study quantified the impact of COVID-19 outbreaks in different stages of the pandemic on passenger traffic at Beijing Daxing Airport and developed a short-term airport passenger traffic forecasting model with selected pandemic variables. This study makes two major scientific contributions. First, the correlation between pandemic trends and daily passenger traffic was examined in the two years after entering the pandemic, covering multiple outbreaks that have occurred in China. This enabled us to compare the impact of outbreaks with different durations and total confirmed cases on passenger traffic. By analyzing patterns of passenger decline and recovery resulting from the early and late outbreaks, we revealed the changes in travel behaviors of passengers and discussed their association with the pandemic control policy, vaccine uptake and virus evolution. Second, a week-ahead passenger traffic forecasting model with historical pandemic variables was developed and compared to a baseline model with only conventional predictors. The predictive power of different pandemic variables was assessed using an explainable artificial intelligence approach called SHapley Additive exPlanations (SHAP) values.

The implications of this study are twofold. First, the correlation between pandemic trends and airport passenger fluctuations was revealed, which is informative for optimizing airport terminal operations to achieve better energy efficiency and IEQ. Second, this study demonstrated a framework of developing a machine learning forecasting model to account for the impact of pandemic situations on airport passenger forecast. This provides an important reference for airport managers to understand and forecast passenger patterns amid the pandemic.

The content of this paper is organized as follows: Section 2 presents an overview of the collected data. Section 3 introduces the adopted statistical methods and machine learning algorithm. In Section 4, we first present the results of the correlation analysis between pandemic variables and passenger traffic, followed by a time series analysis of passenger traffic changes at different stages during the five outbreaks in Beijing, and then present the results of the development, optimization and interpretation of the proposed passenger forecasting models. The impact of pandemic control policies and changes in passenger travel behavior during the COVID-19 pandemic are discussed further in Section 5.

2. Data collection

In this section, the data used to quantify the impact of the COVID-19 pandemic on airport passenger traffic and to build the forecast models are presented.

2.1. Passenger traffic data

The daily passenger traffic at Beijing Daxing Airport from January 1, 2020, to December 30, 2021, was obtained and analyzed. Daxing Airport began operations in September 2019 as the second international airport serving Beijing, in addition to the existing Beijing Capital International Airport. The daily passenger traffic, including arriving and departing passengers for both domestic and international flights, was collected, as shown in Fig. 2 . As the airport began operations in September 2019, passenger traffic was still in the process of growth in early 2020 and had not yet reached its design capacity when the pandemic occurred. Passenger traffic dropped sharply after the first outbreak in January 2020 and did not begin to recover until May 2020. The recovery period after the first outbreak did not last long, as another outbreak in Beijing in June 2020 caused a sharp drop in passenger traffic again (introduced in subsection 2.2). After the end of the second outbreak, passenger traffic experienced a long and steady recovery period, reaching its expected capacity of approximately 100,000 passengers per day in October 2020. However, passenger traffic continued to fluctuate in 2021 due to repeated small-scale outbreaks. There were three noteworthy drops in passenger traffic in January, August, and October 2021. Even so, from 2020 to 2021, annual passenger traffic increased by 56%.

Fig. 2.

Daily passenger traffic at Beijing Daxing Airport.

Notably, the Chinese government has imposed strict restrictions on international flights since February 2020 to control the introduction of COVID-19 cases from abroad. At that time, Daxing Airport suspended international flight services, which did not resume by the end of 2021. Therefore, the correlation between pandemic trends and international flights was not investigated in this study. Only domestic passenger traffic will be discussed and analyzed in the following sections.

2.2. COVID-19 pandemic data

The COVID-19 pandemic data were collected from daily COVID-19 reports released by the National Health Commission (NHC) of the People's Republic of China (http://www.nhc.gov.cn/) and the municipal health boards. Reports from the NHC provided information on daily confirmed cases, daily deaths, daily imported cases, and daily cured cases for 34 provincial-level divisions (e.g., provinces, autonomous regions, municipalities, and special administrative regions) in China. The detailed city-scale pandemic data were released by the city's health commission, for example, the Beijing Municipal Health Commission (http://wjw.beijing.gov.cn/). National and city-scale COVID-19 pandemic data were collected from January 1, 2020, to December 31, 2021, matching the period of the collected passenger traffic data.

An obvious correlation was observed between the pandemic trends in Beijing and the air passenger traffic at Daxing Airport, as shown in Fig. 3 . The red bars represent the new daily confirmed COVID-19 cases in Beijing, and the blue line represents the daily total passenger traffic at Daxing Airport. Five COVID-19 outbreaks that occurred in Beijing in 2020 and 2021 were identified, as highlighted in light red in Fig. 3. We defined a COVID-19 outbreak as with at least 10 confirmed cases in the city. Therefore, those cases that appeared sporadically in between identified outbreaks were not considered. Note that this is not a consensus criterion for defining an outbreak but was based entirely on the pandemic situation in China at the time. The five outbreaks differed greatly in total cases and durations, with the earlier outbreaks being much more severe than the later ones. Statistics of the five outbreaks are presented in Table 1 . The first wave of outbreak lasted the longest and had the highest total number of confirmed cases from January 2020 to April 2020. The second wave lasted only 23 days but developed most rapidly, having the highest number of daily confirmed cases. The following three outbreaks were much milder in terms of the number of total cases and growth rate, with the smallest, the fourth outbreak, having only 13 confirmed cases. However, these mild outbreaks still dramatically reduced passenger traffic.

Fig. 3.

New daily confirmed COVID-19 cases in Beijing and daily total passenger traffic at Beijing Daxing Airport. The five outbreaks in Beijing are highlighted in light red. (For interpretation of the references to color in this figure legend, the reader is referred to the Web version of this article.)

Table 1.

Statistics of five COVID-19 outbreaks that occurred in Beijing in 2020 and 2021.

Outbreak	Duration (days)	Total cases	Mean daily cases	Peak daily cases
1	84	590	7	32
2	23	334	14.6	36
3	41	79	1.9	7
4	20	13	0.7	3
5	33	54	1.6	6

The pandemic trends in the five destination cities with the most connecting flights to Beijing Daxing Airport in 2020 and 2021 were also reviewed, as shown in Fig. 4 . The five cities were Chengdu, Chongqing, Guangzhou, Shenzhen, and Hangzhou. Following the same criterion, we identified five outbreaks that had at least 10 confirmed cases in the five cities, as highlighted in Fig. 4. However, there was less correlation between the number of daily confirmed cases in these cities and passenger traffic at Daxing Airport, except for the first and fourth outbreaks that overlapped with the first and fifth outbreaks in Beijing, respectively. The third outbreak occurred in Guangzhou, with a total number of 146 cases, slightly reducing passenger traffic but by a much smaller amount than the fourth outbreak in Beijing in late July 2021.

Fig. 4.

New daily confirmed COVID-19 cases in five destination cities with the most connecting flights to Beijing Daxing Airport and daily total passenger traffic at Beijing Daxing Airport. The five outbreaks are highlighted in light red. (For interpretation of the references to color in this figure legend, the reader is referred to the Web version of this article.)

In addition, passenger traffic declined slightly in late June 2021, but without simultaneous COVID-19 outbreaks in both Beijing and major destination cities. This was primarily due to military activities and security missions affecting the flight schedule on the eve of the 100th anniversary of the founding of the Communist Party of China in Beijing on July 1, 2021.

3. Methodology

In this section, the statistical methods and machine learning algorithms adopted in this study are introduced. The methodology flowchart is shown in Fig. 5 . First, a correlation analysis was conducted to quantify the impact of various pandemic variables on daily passenger traffic at Beijing Daxing Airport. Then, the impact of multiple COVID-19 outbreaks that occurred in Beijing on airport passengers was characterized based on a time series analysis. Based on the collected data, two LightGBM models were trained, optimized, and compared, i.e., a baseline model with only conventional variables and a proposed model with additional pandemic variables. Finally, the predictions of the proposed model were interpreted using SHAP values and thus the importance of pandemic variables was measured.

Fig. 5.

Research methodology flowchart.

3.1. Correlation analysis

Pearson's correlation, Spearman's rank correlation and Kendall's rank correlation were the most frequently used approaches in previous studies. As a parametric hypothesis test, Pearson's correlation assumes that both variables are normally distributed. However, the Shapiro–Wilk normality test indicated that neither the pandemic variables nor the passenger traffic variable in this study was distributed normally. Spearman's rank correlation and Kendall's rank correlation are both nonparametric methods that do not rely on the assumption of normal distribution, while Kendall's rank correlation is more suitable for handling variables that contain identical values (called “ties”). Because there were many ties in the investigated variables, Kendall's rank correlation was adopted in this study.

In Kendall's rank correlation, for a pair of observations $\left(x_i,y_i\right)$ and $\left(x_j,y_j\right)$ of two variables $x$ and $y$, they are said to be a concordant pair if the elements of one pair are greater than or equal to or less than the corresponding elements of the other pair; otherwise, they are said to be a discordant pair. Thus, Kendall's correlation coefficient ${\tau }_B$ is mathematically defined as:

$$\begin{array}{c}{\tau }_B=\frac{n_c-n_d}{\sqrt{\left(n_0-n_1\right)\left(n_0-n_2\right)}}\end{array}$$ (1)

where $n_c$ is the number of concordant pairs; $n_d$ is the number of discordant pairs; and $n_0$, $n_1$, and $n_2$ are defined as:

$$\begin{array}{c}{n}_0=\frac{n\left(n-1\right)}{2}\end{array}$$ (2)

$$\begin{array}{c}{n}_1=\sum _i\frac{t_i\left(t_i-1\right)}{2}\end{array}$$ (3)

$$\begin{array}{c}{n}_2=\sum _j\frac{u_j\left(u_j-1\right)}{2}\end{array}$$ (4)

where $n$ is the number of values; $t_i$ is the number of tied values in the ith group of ties for the first quantity; and $u_j$ is the number of tied values in the jth group of ties for the second quantity.

Kendall's correlation coefficient has a value between 1 and -1, where 1 represents a perfect positive relationship, 0 represents no relationship and −1 represents a perfect negative relationship. Based on the absolute value of the correlation coefficient, the effect size of the correlation was categorized as small (0.1 ≤ $\left|{\tau }_B\right|$ < 0.3), medium (0.3 ≤ $\left|{\tau }_B\right|$ < 0.5), and large (0.5 ≤ $\left|{\tau }_B\right|$) [39].

3.2. LightGBM model

Machine learning (ML) methods are increasingly being adopted in studies for forecasting air travel demands because they are highly capable of addressing complex nonlinear relationships between variables, as well as their superior computational efficiency in handling massive data [40]. In this study, a tree-based ML algorithm named Light Gradient Boosting Machine (LightGBM) was employed to forecast week-ahead air passenger traffic at Daxing Airport based on historical pandemic data and passenger traffic data. LightGBM is a framework that was developed based on many effective gradient boosting decision trees (GBDTs), such as XGBoost, and has been found to be more accurate and faster in many scenarios. Like other ensemble learning algorithms, LightGBM grows multiple classification and regression trees (CARTs) sequentially based on training data, with subsequent trees being trained with the residuals of preceding trees, as shown in Fig. 6 . LightGBM has two important features that make it efficient and scalable: gradient-based one-side sampling (GOSS) and exclusive feature bundling (EFB) [41]. GOSS enables the algorithm to exclude a significant proportion of training data with small gradients, while EFB enables the algorithm to optimize the number of features by bundling exclusive features. Therefore, LightGBM accelerates the training process of traditional GBDTs by up to 20 times or more, with almost no loss of accuracy.

Fig. 6.

Illustration of the tree growth process in LightGBM.

The performance of the ML model heavily relies on the optimization of model hyperparameters that control its learning process [42]. The conventional approaches for hyperparameter optimization are grid search, which exhaustively evaluates every position of a predefined search space of hyperparameters, and random search, which randomly evaluates the points in a predefined search space. Both methods require a large number of trials to locate the best set of hyperparameters, which is time-consuming. In this study, we used a more effective hyperparameter optimization framework called Optuna [43]. Optuna provided multiple hyperparameter-sampling and pruning algorithms that significantly improved the cost effectiveness of the optimization process. The TPESampler integrated in Optuna was used to sample hyperparameters from the predefined space in this study [44]. In each searching trial, the TPESampler fit two Gaussian mixture models (GMMs), one fitting the set of hyperparameters associated with the best objective value and the other fitting the remaining hyperparameter values. The hyperparameter value that maximized the ratio of the target value of the two GMMs was then chosen. The results of hyperparameter optimization are presented in subsection 4.3.

3.3. SHAP values

Although ML models have demonstrated great success in addressing complex forecasting issues, the interpretability of model prediction has been a critical barrier to adoption in many scenarios [45]. The superior performance of ML models is often achieved by increasing the complexity of the model, which makes it extremely difficult to interpret the output of this “black box”. To address this problem, scientific interest in the field of explainable artificial intelligence (XAI) has grown rapidly, and many approaches have been proposed to help interpret complex ML models [46,47]. An important implication is to understand the effect of the input variables on model prediction in terms of its magnitude and direction [48,49].

In this study, a unified approach called SHAP values was used to interpret the prediction of the resulting LightGBM models [50]. Different from many other approaches that can only provide an overall evaluation of variable importance on the entire dataset, SHAP values allow for a local measure of variable importance, i.e., variable importance is calculated for each observation in the dataset. Specifically, for each input variable, the SHAP values measure the change in the model prediction when conditioning on that variable. Therefore, the SHAP values of a variable can be positive or negative, depending on whether the variable contributes positively or negatively to the model prediction. For each observation, SHAP values attributed to all features collectively explain why model prediction differs from the base value (the mean of the target value in training data). Fig. 7 illustrates an example of the SHAP values for a model with five variables. The SHAP values attributed to variables 1, 2 and 5 are −1.5, −2.0 and −1.5, respectively, indicating that these three features have different degrees of negative impact on the model prediction. Meanwhile, the SHAP values indicate that variable 3 and variable 4 have positive contributions to the model prediction. Considering the effects of all features, the predicted value was lower than the base value by 2.5 in the model. After calculating the SHAP values for all observations, the higher the sum of the absolute SHAP values attributed to a variable, the more important that variable is.

Fig. 7.

An illustration of SHAP values.

3.4. Software and package

Kendall's correlation analysis was performed in R with the base package “stats”. The LightGBM model was established in Python with the package “lightgbm” [41]. The optimization of the LightGBM model was performed in Python with the package “optuna” [43]. The SHAP values were calculated and visualized in Python using the “shap” package [50].

4. Result

4.1. Correlation analysis between pandemic variables and passenger traffic

Kendall's correlation analysis indicated that new daily confirmed cases in Beijing were the most relevant pandemic variables to daily passenger traffic at Daxing Airport, followed by national new daily cases and new daily cases in the five most connected destination cities, as shown in Table 2 . All the correlation coefficients reached a significance level of P < 0.001. Table 2 also shows a slightly higher correlation between the new confirmed cases and arriving passengers compared to departing passenger traffic. The results indicated that during the outbreaks, the number of people flying to Beijing decreased more than the number of people flying out of Beijing. This was partly due to the higher level of restrictions on travel to Beijing than to other Chinese cities during the outbreaks. In addition, the correlation between new daily deaths and passenger traffic was examined but turned out to have a very small effect size.

Table 2.

Correlation analysis of pandemic variables and daily passenger traffic.

Variables	Kendall's correlation coefficient
	Total passengers	Arriving passengers	Departing passengers
Beijing new daily cases	−0.377***	−0.381***	−0.370***
National new daily cases	−0.320***	−0.330***	−0.307***
New daily cases in the five most connected destination cities	−0.213***	−0.212***	−0.214***
Beijing new daily deaths	−0.114***	−0.110***	−0.114***

Significance level: ***, P < 0.001.

A significant hysteresis effect of the pandemic trends on passenger traffic is shown in Fig. 3. Passenger traffic declined rapidly with outbreaks of COVID-19 cases but recovered much more slowly after the outbreaks subsided. Trends in passenger traffic were sometimes determined not by new cases on the same or previous few days but by pandemic trends over a longer period of time before. This hysteresis effect cannot be fully explained by analyzing only the correlation between the number of cases and passenger traffic on a day-by-day basis. Therefore, we calculated the cumulative number of confirmed cases over the previous one to 100 days and its correlation with passenger traffic to account for this hysteresis effect, as shown in Fig. 8 . The cumulative number of cases had a much stronger correlation with passenger traffic than the daily number of cases. The correlation between the cumulative number of cases and passenger traffic had a U-shape, first increasing and then decreasing as the number of days counted increased. The cumulative number of cases in the previous 46 days had the strongest correlation (in absolute value) with passenger traffic, with a coefficient of −0.665, indicating a very large effect size. The results of the correlation analysis provided valuable information for variable selection in forecasting model development.

Fig. 8.

Correlation coefficient between passenger traffic and cumulative number of cases in Beijing.

4.2. Passenger traffic trends during different stages of the outbreaks

In this section, we further investigated the passenger traffic trends during the five outbreaks that occurred in Beijing between 2020 and 2021, leading to a discussion of the difference in the impact of the outbreaks with different scales and durations. Fig. 9 (a) to (e) show the new daily cases (top panel) and daily passengers (bottom panel) during the first to fifth outbreaks in chronological order. Days with new cases reported are marked with red points. We identified three stages for a typical outbreak in terms of passenger flow and highlighted them in Fig. 9: the decline stage in red, the stabilization stage (at a low level) in yellow and the recovery stage in green. In Table 3 , we present the statistics for the five outbreaks, as well as the decline and recovery rates of passenger traffic resulting from each outbreak. The decline and recovery rates in percentage terms were calculated by dividing the decline and recovery rates by the number of passengers before each outbreak. The percentages of the decline and recovery rates are more representative than the absolute rates due to the different levels of passenger traffic prior to the outbreak. Finally, the number of days $D$ needed to recover to preoutbreak levels after each outbreak was calculated using Eq. (5).

$$\begin{array}{c}D=D_d+D_s+\frac{1}{r}\end{array}$$ (5)

where $D_d$ is the number of days in the decline stage; $D_s$ is the number of days in the stabilization stage; and $1/r$ calculates the number of days to recover where $r$ is the percentage rate of recovery.

Fig. 9.

Daily confirmed COVID-19 cases and passenger traffic during the five outbreaks that occurred in Beijing.

Table 3.

Statistics on the decline and recovery stages of passenger traffic resulting from the five COVID-19 outbreaks in Beijing between 2020 and 2021.

Statistics	Outbreaks
	1	2	3	4	5
Total cases	590	334	79	13	54
Duration of decline (day)	25	6	39	12	17
Decline in percentage (%)	96	95	90	81	82
Decline rate (passengers per day)	1832	7814	2276	7811	4668
Decline rate in percentage (% per day)	3.8	15.8	2.3	6.8	4.8
Recovery rate (passengers per day)	530	1171	1997	2107	1170
Recovery rate in percentage (% per day)	1.1	2.4	2	1.8	1.2
Estimated number of days needed to recover to preoutbreak passenger traffic levels	189	64	96	78	105

All five outbreaks resulted in a sharp drop in passenger traffic, dropping between 81% and 96% from the preoutbreak level. Except for the third outbreak, the decline rates of passenger traffic at the beginning of the outbreaks were significantly higher than the recovery rates after the outbreaks. The average percentage decline rate for the five outbreaks was 6.7% per day, while the percentage recovery rate was only 1.7% per day. As a result, it took a longer period (64–189 days) for passenger traffic to recover to the preoutbreak level after each outbreak.

Some differences in the trends of passenger traffic during the early and late outbreaks were observed. The total number of confirmed cases in the first and second outbreaks was significantly higher than that in later outbreaks, which led to a higher percentage decrease in passenger traffic in these two outbreaks. Kendall's correlation analysis verified a significant correlation between the total number of cases in the outbreak and the resulting percentage decrease in passenger traffic (P < 0.05). Another difference was that the earlier outbreaks had a much longer stabilization stage than later outbreaks. In addition to the effect of the outbreak duration, the increasingly early appearance of the recovery stage was an important reason for the shorter stabilization stage in later outbreaks. In the first outbreak, daily passenger traffic began to recover only when no new cases were reported for two consecutive weeks. For the second outbreak, the recovery stage began as soon as no new cases were reported. In the fifth outbreak, passenger traffic entered the recovery stage while sporadic new cases were still being reported. The forward shift in the beginning of the recovery stage reflected a change in passenger travel behaviors during the COVID-19 pandemic, which was further discussed in subsection 5.2.

4.3. Week-ahead passenger traffic forecast model

Based on the collected data, two LightGBM models with different inputs were built to forecast the daily passenger traffic after one week. Model I, as a baseline model, was trained using conventional predictors, including historical passenger traffic variables, date characteristics and quarterly GDP. The quarterly GDP data were collected from the Chinese National Bureau of Statistics (http://www.stats.gov.cn/). In addition to above variables, Model II included historical pandemic variables, i.e., new daily confirmed cases in the previous 7–13 days and cumulative cases within the previous 2/4/6/8/10/12/14 weeks, excluding the week before the forecast date. The names and descriptions of the variables for the two models are shown in Table 4 .

Table 4.

Variables used in Model I and Model II.

Variables	Description	Model I	Model II
N_pre_7/8/9/10/11/12/13	Daily passenger traffic 7/8/9/10/11/12/13 days prior	✓	✓
Month	Month, [1, 2, …12]	✓	✓
Weekday	Day of week, [1, 2, …7]	✓	✓
Holiday	Holiday, [0, 1]	✓	✓
GDP	Quarterly GDP	✓	✓
Case_pre_7/8/9/10/11/12/13	New daily cases 7–13 days prior		✓
Case_pre_sum_2/4/6/8/10/12/14	Cumulative cases within the previous 2/4/6/8/10/12/14 weeks, excluding the week before the forecast date		✓

The collected data was randomly split into training data (60%), validation data (20%) and test data (20%). The validation data were used to optimize the hyperparameters, while the test data were used to evaluate the accuracy of model prediction. Table 5 presents the descriptions and search ranges of optimized hyperparameters, as well as the best set of hyperparameters for the two models located using Optuna [43].

Table 5.

The optimization of hyperparameters for the LightGBM models.

Hyperparameters	Description	Search range	Optimization result
			Model I	Model II
feature_fraction	Percentage of variables used to train each tree model	[0.4, 1.0], float	0.99	0.43
bagging_fraction	Percentage of data used to train each tree model	[0.4, 1.0], float	0.97	0.82
learning_rate	Weighting of new tree model added to the model	[0.01, 0.2], float	0.18	0.18
bagging_freq	Number of iterations to perform bagging	[1,7], integer	6	2
num_leaves	Number of leaves in each tree model	[4,32], integer	18	31
min_child_samples	Minimum number of data needed in a leaf	[20,40], integer	20	21

Table 6 presents the accuracy of the two models on training, validation and test data measured by the CVRMSE (the coefficient of the variation of the root mean square error) and R² (coefficient of determination). The mathematical definitions of the CVRMSE and R² are shown in Eq. (6) and Eq. (7). A lower CVRMSE and higher R² indicated better model accuracy, while a higher CVRMSE and lower R² indicated the opposite. Both metrics showed a consistent result in that the inclusion of pandemic variables significantly improved the models' accuracy on training, validation, and test data. Specifically, compared to Model I, Model II showed a 27.7% decrease in the CVRMSE and an increase in R² from 0.871 to 0.933 on the test data. Additionally, the difference in the performance of Model II on training and test data was smaller than that of Model I, indicating that the inclusion of pandemic variables did not cause overfitting problems.

$$\begin{array}{c}CVRMSE=\frac{1}{\bar{y}}\sqrt{\frac{\sum _{i=0}^{i=n}{\left(y_i-{\hat{y}}_i\right)}^2}{n}}\end{array}$$ (6)

$$\begin{array}{c}{R}^2=1-\frac{\sum _{i=0}^n{\left(y_i-{\hat{y}}_i\right)}^2}{\sum _{i=0}^n{\left(y_i-\bar{y}\right)}^2}\end{array}$$ (7)

where $\bar{y}$ is the mean of the observed values; $y_i$ is the $i^{th}$ observed value; ${\hat{y}}_i$ is the $i^{th}$ predicted value; and $n$ is the number of observations.

Table 6.

Model accuracy on training, validation, and test data.

Model	CVRMSE			R2
	Training	Validation	Test	Training	Validation	Test
Model I	0.114	0.210	0.231	0.972	0.914	0.871
Model II	0.069	0.143	0.167	0.990	0.960	0.933

The SHAP values were calculated to measure the contribution of input variables to model predictions. Fig. 10 shows the SHAP values of the 20 most important input variables. The horizontal coordinate reflects the SHAP values, and the color of the points represents the value of the variables in each observation. The variables are sorted in descending order in vertical coordinates based on the sum of the absolute SHAP values attributed to them. The closest and second closest daily passenger traffic values to the forecast date (N_pre_7 and N_pre_8) were found to have the most contributions to the model, which is not surprising in time series forecasting. The cumulative number of cases within the previous 6 weeks (Case_pre_sum_6) was the third most important variable. Considering that the week before the forecast date was excluded, this result was highly consistent with the result of the correlation analysis that the cumulative number of cases in the past 46 days had the highest correlation with passenger traffic on this day. In addition, the cumulative number of cases within the previous 2/4/10/8 weeks was also included in the ten most important variables. In contrast, the number of daily cases in previous days had much smaller impacts on model predictions, with the number of daily cases seven days earlier (Case_pre_7) being the only one of the ten most important variables. Other conventional factors, such as GDP, month, weekdays, and holidays, also did not have many impacts.

Fig. 10.

Variable importance measured by SHAP values.

5. Discussion

Based on the results, we further discussed the impact of pandemic control policies, changes in passenger travel behaviors and implications of the forecast model to provide additional insight for policy formulation and airport operations.

5.1. Impact of pandemic control policy

The policy and regulations imposed by governments for pandemic control are a crucial factor that influence air travel demands patterns amid the pandemic [32,51]. Most countries around the world had- implemented various restrictions on domestic and international travel to contain the spread of the virus [52], among which China had one of the most stringent policies in force. At the time of research, China was one of the few countries that still adhered to a Zero-COVID policy, which calls for strict containment and closure measures and mass nucleic acid testing when cases emerge, with the goal of eliminating cases within a short period of time. In contrast, many other countries (e.g., the United States) had eased or completely removed their restrictions on domestic and international transport in response to the evolution of the virus, the increasing vaccination uptake rate, and economic considerations.

To explore the impact of different pandemic control policies, the air passenger traffic trends identified in this study were compared with another study that analyzed airports passenger traffic in the United States during 2020 and 2021 [21]. Fig. 11 presented the comparison of passenger recovery ratio between the United States and Beijing Daxing Airport. The ratio is defined as the daily passenger traffic divided by the maximum daily passenger traffic from 2020 to 2021. The passenger numbers in the US were collected by the Transportation Security Administration (data is available at https://www.tsa.gov/coronavirus/passenger-throughput). At the beginning of the COVID-19 pandemic, China and the US went through a similar decline stage, where passenger traffic rapidly decreased to approximately 10% of the max capacity in a few weeks. However, the recovery rate in China was much higher, with passenger traffic returning to approximately 75% of the max capacity within ten months after the first outbreak. In comparison, the passenger traffic at most analyzed airports in the United States did not recover to 50% of the pre-outbreak level after a year of entering the pandemic [21]. On the other hand, China's Zero-COVID policy has also caused greater volatility in air travel demands compared to the US aviation market. The passenger traffic in China decreased significantly whenever new COVID-19 cases arose. An example is that a small outbreak of only 13 cases in Beijing reduced air passenger traffic at Daxing Airport by 81% in less than 2 weeks in August 2021 (see Table 3). Although passenger traffic quickly recovered to preoutbreak levels within two to three months after these mild and short outbreaks, such uncertainty can be a significant challenge to airport and airline operations. In comparison, the US aviation market experienced a more stable recovery, even though the number of confirmed cases was much higher than that in China during the same period.

Fig. 11.

Passenger recovery ratio for the US and Beijing Daxing Airport.

It is not beneficial to claim which pandemic control policies, be they strict or otherwise, are superior to others because multiple trade-offs involving social, cultural, and economic factors need to be considered. Even in terms of the impact on the aviation industry alone, there are short-term and long-term outcomes that need to be weighed. Strict containment policies can quickly quell an outbreak and clear COVID-19 cases in the short term but result in dramatic fluctuations due to recurring cases, while mitigation policies lead to a slower but steadier recovery in air travel demand. Looking ahead, it may be fruitful to explore the possibility of merging different policies to achieve optimization of outcomes in the short and long terms. For example, testing and quarantine regulations can be adjusted in a timely manner to regional pandemic conditions to avoid disproportionate losses due to a few small outbreaks with overly tight regulations. A comprehensive review of the policies adopted by different countries and quantification of their impact would provide more valuable insights to tackle this problem.

5.2. Changes in travel behaviors during early and late outbreaks

We observed several differences in the impact of the five outbreaks that occurred in Beijing in 2020 and 2021 on passenger traffic. The percentage of passenger decline was higher in the two outbreaks in 2020 than in the three outbreaks in 2021. Additionally, late outbreaks had higher recovery rates than the first outbreaks and required fewer days to return to preoutbreak levels. Similar findings were reported in Korea, where the recovery rate of domestic air passenger traffic after the second outbreak was higher than that after the first outbreak [51]. In addition, we observed a forward shift in the beginning of the recovery stages of later outbreaks compared to early outbreaks. Passengers did not begin to increase until there were no new cases reported for two consecutive weeks after the first outbreak but increased when there were still sporadic cases being reported in the later outbreak in 2021.

Changes in travel behaviors were the determining factor for different patterns of passenger decline and recovery between the early and late outbreaks. In the early days of the pandemic, serious concerns about contracting the COVID-19 virus led to a long hesitation period for passengers to resume their usual air travel plans. However, evidence from social media indicated that negative perceptions of COVID-19 changed with the start of the vaccination campaign [53]. Eighty-six percent of China's population is fully vaccinated with the last dose of the primary series, which is among the highest vaccination rates in the world (WHO data, https://covid19.who.int/table). The high vaccination rate might induce a psychological effect that helps people feel less anxious and fearful about the virus and feel reassured to take necessary trips when outbreaks are mild or subsided. The changing characteristics of COVID-19 could be another factor in alleviating passengers' concerns about traveling amid the pandemic. Epidemiological findings suggested that recent variants of COVID-19 were becoming increasingly infectious but caused less severe illness, fewer hospitalizations, and a lower risk of death than the original virus and previous variants [[54], [55], [56]]. In response to the uptake of the vaccine and the constant evolution of the virus, 87% of travelers agreed that COVID-19 will not disappear, so we need to manage its risks while living and traveling normally [57]. This evidence indicated that passengers are adapting to living with COVID-19 in hopes of striking a balance between mobility and pandemic control measures.

5.3. Implications of the forecast model

The results of model comparison showed that including pandemic variables significantly improved the accuracy of week-ahead passenger traffic forecast. The SHAP values further revealed that the cumulative COVID-19 cases in the previous weeks were stronger predictors than daily COVID-19 cases, indicating the hysteresis effect of pandemic trends on passenger traffic. However, in another forecast model research based on the U.S. aviation market, weekly economic index was found to be more important than pandemic variables [29]. Such difference is believed to be resulted by the different pandemic control policy adopted in the two countries. It can be inferred that stricter pandemic control policies would lead to a stronger correlation between pandemic variables and airport passenger traffic, resulting a more significant improvement of model performance after accounting for pandemic conditions. This association dictates that a forecast model based on data from one country will likely not be applicable to another country with a vastly different pandemic control policy. In addition, there are other contextual factors which can moderate the impact of COVID-19 pandemic on passenger travel behavior, such as the vaccination rate, virus variants, social and cultural factors, etc. It is a promising but challenging topic to quantify the impact of these factors and incorporate them into forecast models in an appropriate manner. Consideration of these contextual factors will be beneficial to further improve the performance of the forecast model proposed in this study.

6. Conclusion

The COVID-19 pandemic has caused unprecedented uncertainty in airport passengers, posing a significant challenge to airport terminal operations and management. In this study, we developed and tested a machine learning model for week-ahead passenger traffic forecasting using pandemic-related variables, which outperformed a baseline model using only conventional predictors, such as historical passenger data, time-related inputs, and quarterly GDP. The inclusion of pandemic variables reduced the model error on the test dataset by 27.7%, as measured by the CVRMSE, and increased the R² of the model from 0.871 to 0.933. The SHAP values revealed that the cumulative number of cases in previous weeks contributed more to model prediction than the daily number of cases, indicating the hysteresis effect of pandemic trends on passenger traffic. Meanwhile, conventional predictors such as date characteristics and economic indices had little impact on model predictions under the current pandemic scenarios. These results further affirmed the importance of considering pandemic situations in air travel demand forecasting and provided valuable suggestions for variable selection and model development.

The impact of five COVID-19 outbreaks that occurred in Beijing from 2020 to 2021 on passenger traffic were quantified and compared. The five outbreaks recorded varying number of cases from 13 to 590, but all led to a rapid and dramatic drop in airport passenger traffic, ranging from 81% to 96%. A hysteresis effect of pandemic trends on air travel demand was observed, with the cumulative number of cases in the previous weeks correlating more strongly with passengers than the number of daily cases. As the number of days counted increased, the correlation between the cumulative number of cases and passenger traffic showed a U-shaped change, with the highest absolute correlation coefficient of 0.665 for the cumulative number of cases in the previous 46 days.

We also observed different air travel behavior patterns in early and later outbreaks. The early outbreaks resulted in a longer stabilization stage between the decline and recovery stages compared to later outbreaks. In the first outbreak, the air passenger traffic started to recover after two weeks when no new cases were reported. Contrarily, the air passengers traffic started to recover when there were still sporadic cases being reported in later outbreaks. These differences indicated an important change in people's travel behavior over different stages of the pandemic, i.e., they waited a shorter period of time to resume their regular travel plans once an outbreak subsided. A high vaccination rate, reduced virulence of the virus, and updates to travel restrictions may have assuaged passengers' concerns about air traveling. These findings provide important insight into the airport operations amid the COVID-19 pandemic.

Credit author statement

Hao Tang: Conceptualization, Methodology, Writing Original Draft, Writing - Review & Editing, Juan Yu: Conceptualization, Methodology, Writing Review & Editing, Visualization, Borong Lin: Conceptualization, Writing - Review & Editing, Funding acquisition, Yang Geng: Writing - Review & Editing, Zhe Wang: Writing - Review & Editing, Xi Chen: Data Curation, Li Yang: Data Curation, Tianshu Lin: Data Curation, Feng Xiao: Writing - Review & Editing.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Data availability

The authors do not have permission to share data.