TELKOMNIKA Telecommunication.
Computing.
Electronics and Control Vol.
No.
April 2026, pp.
527O535
ISSN: 1693-6930.
DOI: 10.
12928/TELKOMNIKA.
Hybrid classicalAequantum ensemble learning for real-time flight delay prediction at Tribhuvan International Airport Pavan Khanal1 .
Nanda Bikram Adhikari2 1 Civil Aviation Authority of Nepal.
Kathmandu.
Nepal 2 Department of Electronics and Computer Engineering.
IOE Pulchowk Campus.
Tribhuvan University.
Lalitpur.
Nepal
Article Info
ABSTRACT
Article history:
This study investigates ensemble learning using classical and quantum-inspired models to predict flight delays at Tribhuvan International Airport (TIA).
Nepal.
It combines traditional machine learning algorithms with quantum-based approaches, quantum boosting (QBoos.
and the hybrid QBoostPlus, leveraging quantum properties for faster computation.
The dataset includes flight records from 2020 to 2024 and Meteorological Aerodrome Reports (METAR), analyzed across four sea- sons to capture delay patterns in domestic and international A combined seasonal dataset assesses model generalization.
Six models.
VotingClassifier, adaptive boosting (AdaBoos.
, xtreme gradient boosting (XGBoos.
, categorical boosting (CatBoos.
QBoost, and QBoostPlus are evaluated based on accuracy, precision, recall.
F1 score, area under the curve(AUC), and execution time.
CatBoost achieved high accuracy .
p to 0.
but slower execution .
p to 10,570.
63 m.
QBoostPlus provides competitive AUC scores .
83Ae0.
with faster execution, improving speed by up to 99.
94% and generating predictions in as little as 6.
46 ms.
Although quantum-inspired models have slightly lower accuracy, their computational efficiency and stability show strong potential for real-time flight delay prediction.
This is the first study applying quantum-inspired ensemble learning to Nepalese aviation data, showing promise for regional airports with limited infrastructure.
Received Jun 16, 2025 Revised Dec 4, 2025 Accepted Jan 30, 2026 Keywords:
Categorical boosting Extreme gradient boosting Machine learning Quantum boosting Quantum boosting plus Quantum machine learning Voting classifier This is an open access article under the CC BY-SA license.
Corresponding Author:
Nanda Bikram Adhikari Department of Electronics and Computer Engineering.
IOE.
Pulchowk Campus.
Tribhuvan University Lalitpur 44600.
Nepal Email: adhikari@ioe.
INTRODUCTION
Tribhuvan International Airport (TIA) in Kathmandu.
Nepal, serves as the nationAos primary international gateway, connecting to over 40 global destinations.
Despite its strategic role.
TIA faces operational challenges due to a single sloped runway, absence of an instrument landing system (ILS), and increasing traffic demand.
According to official TIA data, international passenger traffic grew by 9.
29% in 2024, averaging 13,598 passengers per day .
This surge has intensified congestion, delays, and resource limitations, emphasizing the need for intelligent flight delay prediction systems to support efficient airport operations.
Flight delay prediction has been extensively studied using various machine learning (ML) techniques.
Deep learning approaches, such as convolutional neural networkAelong short-term memory (CNN-LSTM) frameworks, have shown promising results in forecasting delays based on historical data .
, .
Hybrid ML models Journal homepage: https://telkomnika.
id/index.
php/TELKOMNIKA ye
ISSN: 1693-6930
combining different algorithms further improve prediction accuracy .
, .
, while ensemble learning methods like gradient boosting and incremental learning effectively capture complex delay patterns .
, .
Additionally, studies leveraging aviation big data have enhanced delay prediction models .
, and investigations into the impact of short-term features have refined model performance .
, .
Flight trajectory prediction has also benefited from hybrid deep learning techniques, improving four-dimensional .
D) trajectory forecasts .
, and spatiotemporal propagation learning has been proposed for network-wide delay prediction .
Recent advancements include transformer architectures for temporal modeling in airport delay prediction .
, .
In parallel, quantum machine learning (QML) techniques are emerging as novel approaches for aerodynamic classification and time series forecasting in aviation.
Quantum support vector machines (QSVM) and data re-uploading quantum methods have demonstrated potential in handling large-scale spatiotemporal data and traffic forecasting .
, .
, opening new avenues for flight delay modeling.
Compared to hybrid models like stacking and bagging .
, .
, quantum boosting plus (QBoostPlu.
integrates quantum-inspired optimization with ensemble fusion, using area under the curve (AUC)-based weighting to improve speed and accuracy without iterative retraining .
Transformer-based ensembles .
, .
have shown high accuracy in flight delay prediction but with heavy computational costs, limiting real-time use in constrained environments.
Recent QML developments .
in transportation and time-series forecasting, such as quantum data re-uploading, offer competitive accuracy and faster convergence over classical models.
Hybrid quantum models like quantum kernel long short-term memory (QK-LSTM) have improved predictive efficiency and reduced computational costs in climate time-series tasks .
Quantum long short-term memory (QLSTM) shows faster convergence and lower test loss than classical LSTM on solar forecasting .
, while quantum sequential recurrent neural network (QSegRNN) achieves comparable or better accuracy with fewer parameters .
These results highlight QMLAos potential to overcome latency and scalability issues in transportation and aviation forecasting.
This study presents the first application of quantum-inspired ensemble learning for flight delay prediction in Nepal, focusing on TIA.
Existing ML models often lack the speed and scalability needed for real-time use in resource-constrained settings.
To address this, we propose QBoostPlus a hybrid framework combining classical ensembles with quantum-inspired optimization to reduce complexity.
Using a multi-season flight and meteorological aerodrome reports (METAR) weather datasets, the model improves both accuracy and efficiency in delay forecasting.
It supports real-time decision-making and is adaptable to other regional airports, advancing smart airport initiatives.
The key contributions of this study include: .
integrating classical ensemble models with quantuminspired optimization for delay prediction.
proposing QBoostPlus for fast and accurate delay prediction suitable for real-time use.
evaluating seasonal and aggregate datasets to assess model generalization.
demonstrating trade-offs between accuracy and execution time to inform hybrid deployment strategies.
METHOD
Dataset and preprocessing This study utilized two primary datasets: the AviBit Traffic Solutions Dataset, which includes 12 flight-related features such as flight number, date, scheduled departure and arrival times, travel time, origin and destination, distance and actual arrival time in the training set, and a test set with the same features except actual arrival time.
The second is the METAR dataset, containing 13 meteorological features from the TIA METAR station, including visibility, sky conditions, temperature, wind, pressure, humidity, and precipitation.
Both datasets were clean, with no missing values or duplicates.
Data preprocessing involved merging the datasets into a single data frame (DataFram.
, synchronizing weather data to coordinated universal time (UTC) and rounding timestamps to the nearest hour.
More than 200,000 communication records were collected from 2020 Key subsets that significantly contribute to the modelAos performance include: seasonal data with 9,522 training and 3,742 test samples, and a combined approach with 3,978 training and 1,610 test samples.
Feature engineering included encoding sky conditions, imputing zero values, removing redundant features, and aligning weather stations with origin and destination airports.
Feature scaling was performed using the standard scaler (StandardScale.
to normalize input variables.
For feature selection, columns with excessive missing data were removed, and the top 14 features were selected based on mutual information (MI) MI measures the degree of dependency between each feature and the target variable, allowing us to prioritize inputs that contribute most to predicting delays.
This approach improves model interpretability by TELKOMNIKA Telecommun Comput El Control.
Vol.
No.
April 2026: 527Ae535 TELKOMNIKA Telecommun Comput El Control ye identifying features with the strongest predictive relationships, offering insights into which flight and weather variables most influence delays.
The final selected features comprised one flight characteristic distance.
origin weather features dew point temperature, precipitation, and few clouds at level 1.
and ten destination weather features dry bulb temperature, dew point temperature, wind speed, wind direction, wind gust, pressure, visibility, precipitation, relative humidity, and scattered clouds at level 1.
We acknowledge that seasonal imbalance in the dataset .
, higher flight volumes in spring and summer compared to winter and autum.
may influence MI scoring, as features dominant in peak seasons could be overemphasized.
To mitigate this, feature selection was performed on both seasonal subsets and the combined dataset to ensure generalization across varying traffic conditions.
Figure 1 illustrates the architecture of the QBoostPlus framework, which integrates quantum-inspired optimization within a lightweight ensemble model to enhance convergence, generalization, and operational Figure 1.
System flow diagram Model building and implementation To comprehensively evaluate predictive performance, we implemented three types of models: classical ML models, the QBoost model, and hybrid approaches.
Classical models In our study, we utilized classical ensemble models including adaptive boosting (AdaBoos.
, extreme gradient boosting (XGBoos.
, categorical boosting (CatBoos.
, and voting classifier (VotingClassifie.
for classification.
AdaBoost sequentially improved performance by focusing on misclassified instances .
, .
XGBoost offered high accuracy and efficiency through gradient boosting with regularization .
, .
CatBoost effectively handled categorical features using ordered boosting .
The VotingClassifier combined predictions from multiple models using hard or soft voting, enhancing overall stability and accuracy .
, .
QBoost model QBoost is a quantum-inspired classification algorithm that reformulates problems into quadratic unconstrained binary optimization (QUBO) format for quantum annealing on quantum processing units (QPU.
Hybrid classicalAequantum ensemble learning for real-time flight delay prediction at .
(Pavan Khana.
ISSN: 1693-6930
Due to limited access to D-Wave hardware, we used the simulated annealing sampler (SimulatedAnnealingSample.
from the dimod library, which emulates quantum annealing on classical hardware while preserving the QUBO framework .
Although it mimics quantum concepts like superposition and entanglement, it lacks true quantum features such as tunneling and large-scale parallelism, limiting scalability.
Nevertheless, this approach allows effective testing of quantum-inspired models for classification and optimization.
Hybrid model Ae QBoostPlus QBoostPlus is a hybrid ensemble classification model that combines multiple weak classifiers using AUC-based weighting.
Instead of relying on a single best model, it evaluates each classifierAos AUC on a validation set and assigns weights through exponential scaling, giving more influence to stronger classifiers.
This weighting strategy aligns with ensemble fusion theory, where model contributions are often scaled by performance metrics to maximize overall predictive power .
, .
Predictions are generated by aggregating the weighted outputs, enhancing both diversity and accuracy.
Unlike traditional boosting.
QBoostPlus avoids iterative training and instead focuses on performance-driven fusion of pre-trained models.
The implementation involves selecting the best classifier based on AUC, optionally adding another, and evaluating the modelAos performance and execution time.
Formal equation of QBoostPlus:
= sign wi A hi .
N = number of weak classifiers, hi .
= prediction .
r decision function outpu.
of Pthe i classifier on input x, wi = weight assigned to the i classifier based on its AUC score .
ormalized so i wi = .
, and yC.
= final predicted label .
, 1 or Oe.
Probability estimation .
sing a sigmoid function with temperature scalin.
P .
= 1 | .
= 1 exp Oe T1 i=1 wi A hi .
T = temperature parameter controlling the softness of probabilities.
Evaluation metrics All models were evaluated using standard classification metrics, including accuracy, precision, recall.
F1-score, and AUC-receiver operating characteristic (ROC), to assess their predictive performance comprehensively.
In addition to these evaluation metrics, execution time was recorded to compare the computational efficiency of classical, quantum, and hybrid models, providing insights into both effectiveness and practicality for real-world applications.
Toolset and system configuration The environment used Visual Studio Code .
Python 3.
x, and libraries such as NumPy, pandas, scikit-learn, and Simulated Annealing from the dimod library.
Experiments were run on a system with an Intel Core i5-1035G1 central processing unit (CPU) .
00 GHz, up to 1.
19 GH.
, 8 GB random access memory (RAM), and Windows 11, which supported both ML and quantum-inspired simulations efficiently.
RESULT AND DISCUSSION
Analysis of combined approach for all seasons The combined approach integrates flight data from all seasons into a single training and testing framework, enabling a holistic assessment of delay patterns.
By aggregating seasonal variations, this approach captures recurring operational characteristics such as airport congestion and systemic inefficiencies while benefiting from a larger and more diverse dataset.
As a result, the models exhibit improved stability and reduced susceptibility to overfitting.
In addition, employing a single unified model simplifies deployment and lowers computational overhead, which is essential for real-time operational use.
TELKOMNIKA Telecommun Comput El Control.
Vol.
No.
April 2026: 527Ae535 TELKOMNIKA Telecommun Comput El Control ye Figure 2 illustrates the AUCAeROC performance of the evaluated models under the combined setting.
XGBoost.
CatBoost, the VotingClassifier, and QBoostPlus demonstrate the strongest discriminative capability, indicating reliable separation between delayed and on-time flights.
In contrast.
AdaBoost and QBoost show comparatively weaker performance, suggesting limited robustness under aggregated seasonal conditions.
Figure 3 presents the relationship between predictive performance and execution time.
QBoostPlus achieves the fastest inference time, substantially outperforming other ensemble models.
Although CatBoost and the VotingClassifier attain comparable predictive accuracy, their significantly higher execution times limit their suitability for latency-sensitive environments.
These results indicate that QBoostPlus provides an effective balance between predictive capability and computational efficiency, making it a strong candidate for real-time flight delay prediction.
Figure 2.
AUC ROC curve of combined approach Figure 3.
Classification performance vs.
execution time of combined approach Hybrid classicalAequantum ensemble learning for real-time flight delay prediction at .
(Pavan Khana.
ISSN: 1693-6930
Enhancing statistical robustness of model evaluation To ensure reliable performance estimation, cross-validation and statistical significance testing were Table 1, summarizes the average accuracy and standard deviation obtained from 5-fold and 10-fold cross-validation, along with paired t-test results.
The models demonstrate consistent generalization, with mean cross-validation accuracies ranging from approximately 85% to 93%.
QBoostPlus achieves the highest average accuracy across both validation settings, while low standard deviations indicate stable performance.
Paired t-test results show no significant differences between 5-fold and 10-fold validation .
ll p-values > 0.
, confirming the reliability of the reported estimates.
These results highlight that QBoostPlus delivers strong predictive performance with efficient computational cost, supporting its suitability for practical deployment.
Table 1.
Model performance with cross-validation and significance testing Model AdaBoost CatBoost XGBoost VotingClassifier QBoost QBoostPlus 5-fold 5-fold standard deviation 10-fold 10-fold standard deviation t-Test value Significance level .
Seperate analysis of each season Flight delays at TIA are strongly influenced by seasonal factors.
Winter fog, spring storms, summer congestion and heat, and autumnal weather transitions introduce distinct operational challenges.
To account for these effects, a season-wise evaluation was conducted to assess context-specific model behavior.
As shown in Table 2, reports the classification performance and execution time of each model across the four seasons.
QBoostPlus consistently demonstrates strong predictive performance, achieving its highest accuracy during the summer season while maintaining competitive results in winter, spring, and autumn.
Importantly, it preserves exceptionally low execution times across all seasonal datasets, highlighting its robustness under varying operational conditions.
Table 2.
Classification performance and execution time of models across different seasons Season
Winter
Spring
Summer
Autumn
Model
AdaBoost
CatBoost
XGBoost
VotingClassifier QBoost
QBoostPlus
AdaBoost
CatBoost
XGBoost
VotingClassifier QBoost
QBoostPlus
AdaBoost
CatBoost
XGBoost
VotingClassifier QBoost
QBoostPlus
AdaBoost
CatBoost
XGBoost
VotingClassifier QBoost
QBoostPlus
AUC
Accuracy F1-score Precision Recall Execution time .
TELKOMNIKA Telecommun Comput El Control.
Vol.
No.
April 2026: 527Ae535 ye TELKOMNIKA Telecommun Comput El Control While CatBoost and the voting classifier occasionally achieve comparable accuracy, their substantially higher inference times reduce their practicality in environments with limited computational resources and strict real-time constraints .
Classical ensemble methods such as AdaBoost show higher precision in certain seasons but suffer from reduced recall, particularly during autumn, indicating sensitivity to class imbalance and temporal variability .
, .
QBoostPlus maintains a balanced trade-off between precision and recall, resulting in stable F1-scores even in challenging seasonal conditions.
This behavior aligns with prior studies reporting the difficulty of delay prediction under imbalanced and temporally heterogeneous data distributions .
, .
Compared with recently proposed transformer-based approaches .
, which offer strong predictive performance at the expense of high computational complexity.
QBoostPlus delivers comparable accuracy with significantly lower latency.
Overall, the seasonal analysis confirms that QBoostPlus effectively adapts to diverse operational contexts while preserving computational efficiency.
Its application to flight delay prediction at TIA represents, to the best of our knowledge, the first use of a quantum-inspired ensemble learning approach in the Nepalese aviation domain.
These results suggest strong potential for broader adoption in infrastructure-constrained airports where scalability and real-time responsiveness are critical .
CONCLUSION
This study explored classical and quantum-inspired ML models for flight delay prediction at TIA, introducing a hybrid framework that balances computational efficiency with predictive accuracy.
The findings highlight the potential of quantum-inspired approaches for time-sensitive aviation tasks, particularly in airports with limited resources.
This work contributes to advancing intelligent, adaptive delay prediction systems tailored to complex airport operations.
Future research should focus on implementing this framework using actual quantum hardware and extending it to other regional airports to enhance scalability and practical utility.
ACKNOWLEDGMENTS
The authors thank Tribhuvan International Airport (TIA) and the Department of Hydrology and Meteorology (DHM).
Nepal, for providing communication and meteorological aerodrome report (METAR) data for this study.
FUNDING INFORMATION
Authors state no funding involved.
AUTHOR CONTRIBUTIONS STATEMENT
This journal uses the Contributor Roles Taxonomy (CRediT) to recognize individual author contributions, reduce authorship disputes, and facilitate collaboration.
Name of Author Pavan Khanal Nanda Bikram Adhikari : Conceptualization : Methodology : Software : Validation : Formal Analysis ue ue ue ue ue ue ue ue ue ue : Investigation : Resources : Data Curation : Writing - Original Draft : Writing - Review & Editing ue ue ue ue ue : Visualization : Supervision : Project Administration : Funding Acquisition CONFLICT OF INTEREST STATEMENT Authors state no conflict of interest.
Hybrid classicalAequantum ensemble learning for real-time flight delay prediction at .
(Pavan Khana.
ISSN: 1693-6930
DATA AVAILABILITY
The data used in this study were obtained from Tribhuvan International Airport (TIA).
Due to policy restrictions, the dataset is not publicly available but may be provided upon reasonable request to the corresponding author, subject to institutional or regulatory approval.
REFERENCES