CBL - Campus del Baix Llobregat

Projecte llegit

Títol: Using Machine Learning to optimize aircraft turn-around procedure

Estudiants que han llegit aquest projecte:

Director/a: ALTMEYER, SEBASTIÁN ANDREAS

Departament: FIS

Títol: Using Machine Learning to optimize aircraft turn-around procedure

Data inici oferta: 20-01-2026     Data finalització oferta: 20-09-2026


Estudis d'assignació del projecte:
    GR ENG SIS TELECOMUN
    GR ENG SIST AEROESP
    GR ENG TELEMÀTICA
Tipus: Individual
 
Lloc de realització: EETAC
 
Paraules clau:
Optimization, time-management, aircraft turn-around procedure, Machine Learning
 
Descripció del contingut i pla d'activitats:
This Bachelor's thesis will focus to analyze the aircraft take-off process from a time management and operational optimization perspective in a real airport such as Barcelona-El prat Airport. The study focusses on identifying and analyzing the different
phases of the aircraft departure process, from initial preparation to take-off clearance and execution.
The methodology is based on direct observation, through systematic collection of time-related data. Based on these data, a process flow analysis is carried out to identify the
critical path, potential bottlenecks, and activities with the greatest impact on delaying the take-off.
Based on the results of the observation, a final proposal will be developed trying to optimize the temporal sequencing and coordination of operations, always considering safety requirements and current regulations.
Additionally, models based on artificial intelligence will be used. A basic model will be trained to compare its outputs with the analytical approach.
 
Overview (resum en anglès):
A crucial operational procedure in commercial aviation is aircraft turnaround, which connects an aircraft's arrival with being ready for the subsequent departure. This study investigates how Total Turnaround Time (TTT) can be decreased while maintaining operational requirements and safety restrictions by combining deterministic scheduling, partial parallelization, and machine learning. For thirteen operations, including taxi-out as an extra step following pushback, a deterministic model was created. To indicate which jobs must be finished in order and which can overlap, full-precedence and partial-overlap relationships were established.
A conservative Baseline plan generated a TTT of 84 minutes. The Baseline Optimized model decreased TTT to 73.5 minutes, a 12.5% improvement, by providing regulated overlaps between deplaning, cleaning, catering, and boarding while maintaining the same activity durations. A synthetic dataset of 5,000 turnaround scenarios was created by altering thirteen activity durations and four overlap thresholds within predetermined ranges in order to investigate a broader range of operational settings. Seventeen input variables were used to train a Random Forest Regressor, which produced an R2 of 0.949, a Mean Absolute Error of 1.85 minutes, and a Root Mean Squared Error of 2.37 minutes.
The deterministic scheduling model was then used to further develop Scenario 928, a promising unseen scenario identified by the model. This resulted in a TTT of 53.01 minutes prior to the last safety modification. The final ML-assisted TTT rose to 56.68 minutes after adding the extra restriction that boarding could only start once fuelling was finished. The final ML-assisted configuration was 17.32 minutes, or around 23.4%, shorter than the Baseline Optimized timetable, which achieved 74.0 minutes under the identical safety conditions. Based on 1,000 Monte Carlo simulations with ±10% variance in activity durations, a robustness analysis yielded a mean TTT of 56.73 minutes and a 95th percentile of 59.47 minutes.
The findings demonstrate that turnaround performance is influenced by precedence relationships, overlap potential, and the location of each activity inside the scheduling network in addition to individual task lengths. Therefore, although recognizing that validation with actual operational data would be required before practical deployment, the study supports the use of deterministic scheduling and machine learning as complimentary techniques for analysing turnaround improvement.

© CBLTIC Campus del Baix Llobregat - UPC