Abstract

Introduction

Safety is of utmost importance in aviation. For this reason, pilots and air traffic controllers can abort an approach if a safe landing cannot be continued for a number of reasons [Campbell et al. 2022; Campbell et al. 2018; Coker 2014; Zaal et al. 2019], such as adverse weather conditions [Donavalli et al. 2017], approaches flown “too fast, too low, and too close” [Gluck et al. 2019], unstable approaches [Civil Air Navigation Services Organisation 2013], excessive workload, etc. In these cases, pilots are required to fly a Go-Around (GoA). This is a normal flight procedure in which the pilots abort an approach, climb to a safe altitude, reposition their aircraft, and then make another landing attempt. For instrument approaches, the flight path that aircraft must follow during a GoA is specified precisely in the form of a so-called missed approach procedure.

GoAs occur relatively rarely. As shown later in this study, the frequency of GoA varies between 1.5 and 6 events per thousand approaches flown. Because of their rarity, GoAs are considered safety relevant and require appropriate training of both air traffic controllers and the cockpit crew [De Voogt et al. 2022; Dehais et al. 2017]. Besides that, GoA can have a significant impact on the sustainability of a flight (e.g., due to reduced fuel efficiency as well as increased CO2 and noise emissions [Casado et al. 2023; Murrieta-Mendoza and Botez 2016; Dancila et al. 2014]), the operational performance of an airport or an airline (e.g., reduction in airport capacity [Casado et al. 2021], introduction of delays in flight schedules), as well as the perceived passenger experience.

The literature on GoA is mainly concerned with (i) the identification of factors that favour GoA, (ii) the detection and labelling of GoA events given trajectory data, (iii) the generation of artificial GoA trajectories, and (iv) the prediction of GoA events:

To address the question of which factors can lead to a GoA, Gariel et al. [2011] evaluated a four-year data set of San Francisco International Airport using statistical analysis, while Dai et al. [2019; Dai et al. 2021] applied a principal component logistic regression model to a six-month data set on New York John F. Kennedy Airport. It has been found that a variety of factors, such as adverse weather conditions (poor visibility, low ceiling), increased traffic density, or aircraft traffic on the ground of an airport may affect the occurrence of GoA events.

To detect and label GoA events, data sets containing trajectories of approaching aircraft are required. In the literature, data originating from secondary radar [Gariel et al. 2011] or from Automatic Dependent Surveillance-Broadcast (ADS-B) [Xu et al. 2021; Proud 2020; Figuet et al. 2020] are mainly used for this purpose. For the actual process of detection and labelling, rule-based algorithms are applied [Proud 2020; Dai et al. 2021]. The detection of GoA events on the basis of trajectory data is computationally expensive. Therefore, Monstein et al. [2022] made a data set of trajectories sourced from the Opensky Network (OSN) [Schäfer et al. 2014] available to the public, which contains more than 9 million trajectories labelled as landings and over 33,000 identified as GoA.

As GoAs are rare events, Krauth et al. [2021] developed a method based on multivariate density models to generate artificial GoA trajectories. Subsequently, Krauth et al. [2021] applied this method in a Monte Carlo simulation-based collision risk model to estimate the probability of interaction between aircraft taking off at Zurich Airport and aircraft flying a GoA.

The prediction of GoA events is carried out both on a macroscopic as well as on a microscopic level. The macroscopic level addresses the prediction of GoA events for an airport system. In this regard, Chou et al. [2021] evaluated 18 different supervised machine learning models to investigate which method is best suited to predict GoA occurrences at Denver Airport. It turned out that a CatBoost model, which was fine-tuned for usage at the airport in question, yielded the best results. Moreover, Figuet et al. [2020] proposed the use of a generalised additive model for the prediction of the GoA rate at Zurich Airport. The model considers a number of inputs, such as weather-related factors, traffic density information, aircraft and airline mix, etc., to estimate the probability of GoA in the next hour. In contrast, microscopic prediction models deal with individual flights or trajectories. Gariel et al. [2011] developed a preliminary alert system for GoA prediction based on linear discriminant analysis. Figuet et al. [2020] evaluated seven different machine learning classifiers to evaluate which GoA of aircraft approaching Zurich Airport can be predicted best. As such, a random forest classifier was found to perform best, albeit with a high false positive rate. Similarly, Puranik et al. [2020] employed a random forest model for the online prediction of aircraft landing parameters (e.g., landing speed) and ultimately the probability of GoA. Finally, Dhief et al. [2022] proposed a microscopic GoA prediction model based on the CatBoost and XGBoost algorithms that are employed to predict GoA at Philadelphia and Van Nuys Airports.

While the literature on GoA is rather extensive, there is a notable lack of comprehensive macroscopic studies on GoA’s impact on incoming airport traffic. Addressing this gap, our paper provides a macroscopic examination of GoA events at 20 major European airports. We focus on statistically analyzing the GoA rate and assessing the influence of GoA occurrences on the efficiency of nearby landing traffic. For this purpose, we have compiled an extensive ADS-B data set from the OSN, encompassing over 6.7 million arrival trajectories at these airports from January 2019 to June 2023. Within this data set, we identified more than 20,196 individual GoAs.

Data sets and Methodology

In this study, data originating from two different sources are merged. The first data set contains ADS-B trajectories sourced from the OSN [Schäfer et al. 2014], while the second data set was extracted from the EUROCONTROL Network Manager (EU-NM). Both data sets are introduced below in more detail.

ADS-B trajectories

For this study, we collected ADS-B data of landing trajectories from the OSN at 20 major European airports, covering a time period from January 2019 to July 2023. These 20 airports were selected from the 30 largest European airports in terms of movements. Table 1 summarises the number and distribution of landings that the OSN data set used in this study contains per airport.

Filtering.

As the focus of this study lies on identifying GoA executed by aircraft operated in commercial aviation, we refined our data set by exclusively retaining airplanes equipped with turboprop or jet engines, as inferred from their ICAO aircraft type descriptions. Following this filtering process, our landing data set comprises of approximately 6.7 million flight trajectories.

GoA identification.

Following Monstein et al. [2022], GoAs were identified using the traffic Python library [Olive 2019]: First, each segment of a trajectory is assigned to a specific flight phase, employing the algorithms introduced by Sun et al. [2016]. Then, portions of the trajectory that align with a runway at an airport for at least one minute are identified and, if two distinct runway-aligned segments are separated by a single climb phase, the trajectory is classified as a GoA. Having applied this classification algorithm to all 6.7 million flight trajectories mentioned above, 20,196 flights were identified as performing a GoA.

EUROCONTROL EU-NM

In its role as Network Manager, EUROCONTROL collects detailed data on flights, encompassing both planned and actual occurrences. In this study, two different data sets curated by EUROCONTROL are used to augment the OSN-based landings data set, namely market segment information and Arrival Sequencing and Metering Area (ASMA) data.

Market segment.

Based on EUROCONTROL EU-NM data, a market segment can be assigned to an aircraft movement on the basis of the aircraft type, the aircraft operator, and the ICAO flight type. Thereby, EUROCONTROL distinguishes between seven different market segments, namely mainline, low-cost, business, regional, cargo, non-scheduled, and military.

Arrival Sequencing and Metering Area.

An Arrival Sequencing and Metering Area (ASMA) is a virtual cylinder with radius 40 NM, whose center is located on the reference point of an airport. By using EUROCONTROL’s Correlated Position Report data, which refers to data concerning the position of aircraft derived from Air Traffic Control (ATC) surveillance systems, the entrance of flights into the ASMA of the 20 major European airports considered in this study is detected. More precisely, the time of entry as well as the entry point, i.e., latitude and longitude, of each aircraft is recorded.

Merged data set

In a first step, metadata on the OSN-sourced ADS-B trajectories is extracted. Subsequently, this metadata is merged with its corresponding entries logged in the EUROCONTROL EU-NM data set on the basis of the day of flight, the destination airport, and the transponder address of the aircraft, i.e., the ICAO24 address. In the course of this, trajectories landing at Istanbul Sabiha Gökçen Airport could not be mapped, as this airport is not covered in the EU-NM data set. For all the other airports, more than 95% of the landings could be mapped with the EU-NM data set, with the exception of Manchester Airport for which only 51% of the trajectories could be mapped. The column names of the merged data set, the data types used, as well as a description of all columns are specified in Table 2.

Methodology

The methods section is divided into two parts. First we show how the GoA statistics were compiled, then how the impact of a GoA on arriving traffic is determined.

GoA statistics

Initially, we created GoA statistics by calculating the GoA rate based of the landing airport, the market segment, and the aircraft type. To this end, we first divided the number of observed GoA incidents within each group by the total number of landings within that same group. Subsequently, we calculated the 95% confidence interval for each estimated GoA rate by employing the Clopper-Pearson interval method [Clopper and Pearson 1934].

In a second step, we shifted our focus to the analysis of GoA segments. For this matter, we extracted the portions of flight trajectories that correspond to GoA. By this we refer to the part of a flight from the moment the aircraft crosses the landing runway threshold until it crosses the threshold for its next landing attempt. After identifying the GoA portions of all trajectories containing a GoA, we calculated the duration and distance of each GoA segment and aggregate statistics on a per-airport basis.

In a last step, we estimated the fuel consumption of aircraft performing a GoA. In this regard, we focused exemplary on GoA flown by aircraft of type Airbus A320. To calculate fuel consumption, we first extracted the flight portion between GoA initialization (i.e., the lowest altitude observed during the first approach) and touch-down. Then, we employed the fuel flow model implemented in the OpenAP library [Sun et al. 2020]. This model allows us to estimate fuel consumption using actual flight trajectory data (namely the altitude, the vertical rate, and the speed), with the caveat that it requires information on the aircraft’s mass, which is not available in the ADS-B data. For this reason, we assumed a fixed mass value of 60 tons, i.e., 93% of the maximum landing weight, for each A320 when the GoA is initialized.

Impact of GoA on the arrival traffic

To investigate the impact of GoAs on other traffic arriving at an airport, we developed a methodology based on the ASMA timing method. As such, the ASMA timing method quantifies the actual duration of a flight between its last entry into the ASMA cylinder, see Section 2.2.0.2, and its landing time. With this methodology, we determined an unimpeded ASMA time that reflects the transit time in uncongested airspace conditions. To this end, the unimpeded time is statistically determined by calculating the \(20^{\text{th}}\)-percentile of the observed ASMA times of distinct groups of flights that share common parameters, including aircraft class, ASMA entry sector, and arrival runway. Subsequently, the additional ASMA time is determined for a given flight by calculating the difference between a flight’s actual ASMA time and the corresponding unimpeded ASMA time.

The methodology introduced above is applied on a per runway basis. After retrieving all landings for a runway of interest, we retrieved the bearing between the airport and the aircraft at the moment it enters the ASMA cylinder. We then fitted the distribution of these bearings with a Gaussian kernel density estimation (KDE). The KDE approach allows us to smooth the data and estimate the underlying distribution. Subsequently, we extracted the modes, or peaks, of this distribution by identifying the local maxima of the fitted KDE. Following this, each individual landing is assigned to the nearest mode of the extracted bearings. If the first and last modes are separated by less than 15 degrees, they are considered as one to circumvent the 0 to 360 degrees discontinuity.

With the ASMA sectors identified, we calculated the unimpeded ASMA time for each combination of aircraft class and entry sector by determining the \(20^{\text{th}}\)-percentile of all ASMA times for each combination. To illustrate our methodology, we present an example for runway 25L at Frankfurt Airport in Figure 4. Subsequently, we calculated the additional ASMA time for each landing on this runway.

To assess the impact of GoA on landings at an airport, we compared the additional ASMA times before and after a GoA. Our methodology, applied to every GoA occurrence, starts with the following steps:

1. Baseline - Pre-GoA Additional ASMA Time: We established a baseline by computing the median additional ASMA time for landings within the 20-minute interval preceding a GoA.

2. Post-GoA Additional ASMA Time: We calculated the median additional ASMA times for successive 10-minute intervals following the GoA, ranging from [0, 10min], [1, 11min], up to [50, 60min].

3. Go-Around Arrival Induced Delay: The Go-Around Arrival Induced Delay (GoAID) for each 10min window was determined by subtracting the pre-GoA median from the respective post-GoA median additional ASMA times.

After calculating the GoAID for each time window of every GoA occurrence, we aggregated these values at an airport level. This entailed determining the median GoAID for each interval across all GoA occurrences at the airport. For example, the median GoAID for the [0, 10min] period provides insight into the typical delay experienced during this interval following a GoA at that airport.

To verify whether the GoAID times are directly related to GoA events, we employed a control group approach. Specifically, we applied our analysis methodology to landing sequences that have not been affected by any GoA occurrence. To create these control groups, we randomly selected time periods at each airport during which no GoA occurrences were observed. Ten control groups for each airport were created. This comparative analysis between GoA and non-GoA periods allowed us to discern whether there exists a significant difference in GoAID times. If the GoAID times exhibit notable distinctions during GoA periods as opposed to non-GoA periods, it strongly suggests that GoAs have a discernible impact on landing delays.

Results

This section first presents a series of GoA-related statistics. Then, ASMA timings are used to show how GoAs influence the efficiency of an airport and its arriving traffic.

Statistics

Statistical information on the observed GoAs is provided below. This includes information on the rate of GoA, the distance and duration of the GoA, and the fuel consumption of Airbus A320 aircraft during GoA.

GoA Rates.

The observed GoA rate per airport is illustrated in Figure 6. As such, the GoA rate varies from slightly below 1.5 GoA per 1,000 landings for Rome Fiumicino Airport up to a rate of 6 for Sabiha Gökçen Airport in Istanbul, with most airports recording a value below 3.8 GoA per 1,000 landings. On average, the 20 large European airports analysed in this study show a GoA rate of 3.1 per 1,000 landings.

Figure 7 contains the observed GoA rate for different market segments over the period from January 2019 to July 2023. The figure indicates on the one hand that the Non-Scheduled and Regional market segments recorded a lower GoA rate of approximately 2.2 GoA per 1,000 landings. On the other hand, the Mainline, Cargo, and Low-Cost segments exhibited a higher rate of approximately 3.2 GoA per 1,000 landings. The Business Aviation segment occupied an intermediate position, with a GoA rate of about 2.6 per 1,000 landings.

Figure 8 illustrates the GoA rate according to different types of aircraft. A clear variation is noticeable in the GoA rate across numerous airframe typecodes, with rates ranging from 1 to 4.5 GoA per 1,000 landings.

GoA distances and duration.

Table 3 summarizes the different quartiles for GoA execution time and distance at different airports. Median times vary from 11.5 minutes for London Stansted Airport and 16.5 minutes for Madrid Barajas Airport, while the median distances are ranging from 36.4 NM to 58.2 NM. It is interesting to note that some GoA are particularly long. For example, the longest GoA contained in the data set used in this study was executed at Paris Charles de Gaulle Airport. It lasted for more than 2 hours and covered a distance of 376 NM, as illustrated in Figure 9.

GoA fuel consumption.

Figure 10 contains a box plot showing the distribution of the fuel consumption of the 2,417 GoA flown by aircraft of type Airbus A320 contained in the data set used in this study. These A320-related GoA are analyzed on a per airport-basis. As such, the fuel consumption of an Airbus A320 during a GoA typically ranges between 300 and 600 kg, depending on the airport and on the length/duration of the GoA.

GoA impact on the arrival traffic

In Figure 11, the impact of GoA on the arriving traffic of an airport is quantified by comparing the ASMA timings before and after GoA events. We can observe considerable differences between when a GoA is performed and the matched control groups, where no GoAs were executed. Indeed, the red curve (with GoA) is systematically different from the blue curves (control groups). It is noticeable that most airports evaluated in this study demonstrate a similar trend, with the additional ASMA time reaching a peak value within the 5-25 minute period that follows a GoA occurrence. For instance, at airports such as Milan Malpensa and Dublin, the GoAID peaks at approximately 100 seconds. This indicates that, on average, landings occurring between 10 and 20 minutes after a GoA spend an additional 100 seconds in the ASMA cylinder compared to those before the GoA. Conversely, at some airports, the GoAID is consistently lower, remaining below 50 seconds throughout the hour-long observation period. Notably, Madrid Barajas Airport is the least affected, with a maximum GoAID of 21 seconds. Furthermore, it is noticeable that the recovery speed, which specifies the time until the GoAID returns to pre-GoA levels (i.e., a value of 0) after a GoA occurrence, varies among the airports. At some airports, a recovery occurs within approximately 30 minutes, while at others, the GoAID does not completely return to zero within the hour under consideration after a GoA.

Discussion

This study examined GoA rates at 20 large European airports on the basis of OSN trajectory data. The GoA rates shown in Section 3.1 exhibit large variations with respect to the airports, aircraft types, and market segments investigated. Specific airport examples, such as Istanbul’s Sabiha Gökçen Airport, London Heathrow, and Paris Charles de Gaulle, were identified as the airports with the highest GoA rates, highlighting the role of operational procedures and regulatory implementations, e.g., Simultaneous Independent Parallel Approaches performed at Istanbul’s Sabiha Gökçen Airport, Time Based Separation at London Heathrow, or reduced separation minima on final approach at Paris Charles de Gaulle, which could be further analysed for their impact on GoA rates. Moreover, the correlation of GoA rates with aircraft types was explored. Factors such as in-homogeneous approach speeds, climb rates, energy management, and safety procedures for varying aircraft types, could be possible causes of the observed GoA rate variations. Finally, factors such as operational protocols, economic pressure, and scheduling flexibility could contribute to variations in GoA rates across different market segments. For instance, mainline and cargo carriers, operating larger aircraft, might have higher GoA rates due to stricter safety standards. It can thus be stated that the interplay of all these factors makes the causality analysis for GoA extremely complex.

The GoA execution time and distance across different airports were analyzed. Factors such as airport layout, ATC efficiency, surrounding terrain, prevailing weather conditions, operational procedures, and local regulations could possibly influence GoA execution times, as exemplified by the comparison between Madrid Barajas Airport and London Stansted Airport. Longer GoA execution times and distances could have multiple impacts, including increased fuel consumption as shown for example for aircraft of type Airbus A320 in Section 3.1, emissions, noise pollution, operational delays, ATC workload, passenger inconvenience, safety risks, and financial costs to airlines. These findings highlight the importance of understanding and mitigating GoA impacts to enhance airport operations and aviation sustainability.

The observed increase in ASMA timings following GoA occurrences indicates the necessity for additional time to safely sequence incoming traffic. The observed increase in ASMA timings could stem from various factors, such as the need for expanded spacing, increased ATC workload, and the re-sequencing of arriving flights. Subsequently, these enhanced requirements clearly reflect the operational challenges faced in managing such events, suggesting the necessity for strategies to augment arrival sequencing productivity post GoA occurrences. Future research could delve into the exact reasons for ASMA time variations among different airports. An in-depth exploration of relevant variables, such as runway configuration, air traffic density, and the types of involved aircraft may offer valuable insights.

This study has a number of limitations. The method used in this study to detect GoA based on trajectories of landing aircraft is not perfect. On the one hand, the method cannot detect GoA of flights that diverted to another airport for the final landing. On the other hand, flights that do not align long enough with the extended runway axis may not be detected as GoA. Furthermore, this study is based on a data set, which includes trajectories observed in the period from January 2019 to July 2023. Because this data set contains the period in which international air traffic was affected by the COVID19 pandemic, it must be assumed that the results of this study are influenced by these traffic fluctuations.

Conclusion and Outlook

This study investigated the GoA rates at 20 large European airports based on OSN-sourced aircraft trajectory data. For this purpose, a data set consisting of 6.7 million trajectories of aircraft landing at these airports in the period of January 2019 and July 2023 was compiled. A rule-based algorithm was then used to identify those trajectories in the data set that contained one or more GoAs. A comprehensive examination of the resulting 20,196 trajectories that involve a GoA revealed large differences in terms of GoA rates between different airports, market segments, and aircraft types. Besides, the analysis shed light on the operational challenges posed by GoA execution, resulting in increased fuel consumption. The observation of inflated ASMA timings post-GoA events encapsulates the additional time requisites to ensure safe arrival sequencing, underlining the operational hurdles in managing such unforeseen events within a tightly coordinated arrival sequence.

Our findings highlight the interconnected nature of aviation operations, illustrating how a solitary GoA maneuver extends its influence beyond the individual aircraft, altering the landing sequence of subsequent flights. This understanding becomes essential in the light of recent optimization efforts aimed at enhancing runway capacity and fuel efficiency, such as approaches with steeper glideslopes and reduced separation of aircraft on final approach, especially if this leads to higher GoA rates. The quest for optimizing airport operations while ensuring paramount safety and minimizing environmental footprint necessitates a continued exploration and understanding of GoA dynamics. For this reason, several extensions to this research are possible, such as quantifying the expected total delay one single GoA adds to an airport system. Subsequently, the insights from this study pave the way for informed decision-making and strategic planning towards achieving these overarching goals in the aviation sector.