Queue up for Takeoff: A Transferable Deep Learning Framework for Flight Delay Prediction
Nnamdi Daniel Aghanya, Ta Duong Vu, Amaëlle Diop, Charlotte Deville, Nour Imane Kerroumi, Dr. Irene Moulitsas, Dr. Jun Li, Dr. Desmond Bisandu · 2025
Read Queue up for Takeoff: A Transferable Deep Learning Framework for Flight Delay Prediction by Nnamdi Daniel Aghanya, Ta Duong Vu, Amaëlle Diop, Charlotte Deville, Nour Imane Kerroumi, Dr. Irene Moulitsas, Dr. Jun Li, Dr. Desmond Bisandu (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-09084.txt.
Introduction
Nnamdi Aghanya and co-authors build a flight delay forecast tool in a 2025 paper. Delays cost airlines money and strain travelers and crews. A good tool must stay sharp when moved to a new flight network.
Their model joins queue theory with a light attention net. On United States flight data it beat past tools on both hit rate and balance scores. On a European data set, with no retraining named, it still scored well. That cross network strength is the core claim. Airlines could use such forecasts to plan gates and crews, and to warn travelers sooner.
Catalog record
| CREATOR | Nnamdi Daniel Aghanya, Ta Duong Vu, Amaëlle Diop, Charlotte Deville, Nour Imane Kerroumi, Dr. Irene Moulitsas, Dr. Jun Li, Dr. Desmond Bisandu |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 49 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2507-09084
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Queue up for Takeoff: A Transferable Deep Learning Framework for Flight Delay Prediction Nnamdi Daniel Aghanya, Ta Duong Vu, Amaëlle Diop, Charlotte Deville, Nour Imane Kerroumi, Dr. Irene Moulitsas, Dr. Jun Li, Dr. Desmond Bisandu Abstract Air travel facilitates the movement of people and goods globally. However, flight delays remain a significant issue, profoundly causing financial losses, operational inefficiencies, and passenger dissatisfaction. Delay prediction remains a critical challenge in the aviation industry; and thus, several approaches have been developed to model and predict flight time delays in recent decades. Therefore, in order to provide great passenger experiences, boost revenue, and reduce unnecessary revenue loss, flight delays most be precisely predicted and such models can be generalised across networks. In this paper, a novel approach was developed that combines Queue-Theory and simple attention model, referred to as the Queue-Theory SimAM (QT-SimAM) model to predict flight delays. To validate our proposed model, we used data from the US Bureau of Transportation of Statistics. Our experiment results show that the proposed QT-SimAM (Bidirectional) out-performed the existing methods with accuracy of 0.927, precision of 0.946, recall of 0.927 and F1 score of 0.932. For transferability, we tested our model using the EUROCONTROL dataset. Our experiment results show that the proposed QT-SimAM (Bidirectional) has accuracy of 0.826, precision of 0.794, recall of 0.826 and F1 score of 0.791. Ultimately, the paper outlines an end-to-end methodology for predicting flight delays. Our suggested method’s effectiveness in forecasting flight delays with the fewest possible errors is clearly defined by all evaluation parameters. Additionally, passengers’ anxiety can be successfully reduced by using our robust model’s prediction result to obtain information about the delayed flight ahead of time from aviation decision systems across networks. K eywords Flight delay \cdot Queue-Theory \cdot Simple attention model \cdot Airport \cdot Aviation. 1 Introduction Flight delays remain a significant issue in the aviation industry, profoundly causing financial losses, operational inefficiencies, and passenger dissatisfaction [ 1 , 2 , 3 ] . Predicting flight delays remains a very difficult challenge due to their inherent, mostly inevitable uncertainty arising from a host of factors, such as weather conditions, air-traffic congestion effects, technical problems, airspace capacity, and airline-level operational constraints [ 4 , 5 , 6 ] . Flight delays can be defined as departures that leave more than 15 minutes after their scheduled time or aeroplane arrivals that land over 15 minutes late, a benchmark set by the United States (US.) Department of Transportation and widely adopted in the literature [ 7 , 8 , 9 , 10 , 11 ] . Airline delays result in significant financial and operational losses, costing the global industry billions annually [ 12 , 13 , 14 ] . A delay in one flight often disrupts subsequent flights due to the interconnected nature of airline schedules, affecting aircraft utilisation and crew assignments [ 15 ] . Airlines incur additional costs from increased fuel consumption, crew hours, and passenger compensation [ 16 ] . Furthermore, delays have an environmental impact due to increased carbon emissions from extended taxiing or holding patterns [ 17 ] . Delay prediction has been a critical challenge for decades, with several approaches developed to model and predict flight time delays [ 18 , 19 , 20 , 21 ] . Fortunately, the modern aviation ecosystem generates vast amounts of data [ 22 ] , and researchers have increasingly turned to advanced computational approaches like Machine Learning (ML) and Deep Learning (DL) to harness it [ 23 ] . These predictive models offer the potential to accurately forecast flight delays, enabling proactive decision-making [ 6 ] . However, results can be limited by the absence of specific airline data [ 24 ] , and despite extensive research, mitigating flight delays remains a primary problem to be solved [ 25 ] . A significant gap in existing research is the reliance on weather-delay features for forecasting. This affects the model’s utility for most open datasets that lack such features (e.g., the EU dataset), rendering a universally applicable solution impractical and highlighting the need for a transferable approach. In response, this research predicts flight delays by developing a model that combines the Queue-Theory and simple attention model, referred to as the Queue-Theory SimAM (QT-SimAM) model. This architecture innovatively modifies the SimAM attention mechanism by biasing its energy function with proxies for an aircraft’s accumulated workload, calculated using principles from Queue Theory. This integration allows the model to adaptively increase attention to feature channels associated with congested flight chains, providing a more informed representation of how delays cascade from one flight leg to the next. By addressing some key limitations in existing methodologies, such as the lack of generalizability and limited understanding of delay propagation, the study offers a robust framework for more accurate and transferable flight delay forecasting. The developed models provide technical information for airlines, airports, and air traffic management, demonstrating the potential for a broader application in different aviation networks (transferability). In addition, the research highlights the broader implications of advanced predictive technologies, offering a path to improved operational efficiency, passenger experience, and environmental sustainability in the aviation industry. The rest of the paper is organised as follows. Chapter 2 establishes the flight delay methodology based on the Queue-Theory SimAM (QT-SimAM) model. In Chapter 3, we present prediction results and discussions, while Chapter 4 presents conclusions with limitations. The Appendix Chapter summarises related literature on flight time delays. 2 Methodology This chapter details the methodology employed to develop and evaluate an attention-based framework for predicting flight delays, with a focus on assessing model transferability between the United States (US) and the European Union (EU). The core of this framework is a novel integration of queue-theoretic principles with attention mechanisms. The subsequent sections describe the data acquisition and feature harmonisation steps, followed by the construction of flight chains for sequential modelling. The architectural details of the attention-based baseline models are then presented. The chapter culminates in the introduction of our proposed Queue-Theory Simple Attention Module (QT-SimAM) and its pairing with a Queue Mogrifier (QMogrifier) LSTM. An overview of the end-to-end workflow is provided in Figure 1 . Figure 1: Flight Delay Prediction Pipeline Overview. 2.1 Data Acquisition and Feature Harmonisation This study utilises flight operations data from the US Bureau of Transportation Statistics and EUROCONTROL. Data from March, June, September, and December 2022 was selected to cover diverse operational and meteorological seasons, providing a robust baseline for evaluating model generalisation and transferability between the US and EU. Schema alignment The raw US and EU data files differ in scope and nomenclature. A notable asymmetry is a weather_delay field in the US data with no counterpart in the European feed. A strictly comparable baseline, therefore, rests on the intersection of the two schemas as defined by Equation ( 1 ), \Sigma\;=\;\mathrm{Cols}\bigl(\widetilde{\mathcal{D}}^{\mathrm{US}}\bigr)\;\cap\;\mathrm{Cols}\bigl(\widetilde{\mathcal{D}}^{\mathrm{EU}}\bigr), (1) which yields 38 shared features. To measure the incremental value of local features, an auxiliary set of US-only models is also trained with the weather_delay feature reinstated. Minor gaps were repaired by imputation: zeros for continuous variables and sentinel tokens for categoricals. No rows were dropped, avoiding region-specific sampling bias. The final harmonised datasets comprised approximately N_{\text{US}}=384,211 (US) and N_{\text{EU}}=327,408 (EU) rows. Feature catalogue In table 1