Campus del Baix Llobregat
 
Projectes oferts

Projectes matriculats

Tribunals i dates de lectura

Projectes llegits
(2019-2)

DG ENG AERO/SIS TEL

DG ENG AERO/TELEMÀT

DG ENG SISTE/TELEMÀT

ENG TELEC 2NCICLE 01

ET AERO/ETT SIST 05

ET AERONÀUTICA 03

ETT SIST/ ET AERO 05

ETT SIST TELEC 00

ETT SIST TELEC 91

ETT TELEMÀTICA 00

GR ENG AERONAVEGACIÓ

GR ENG AEROPORTS

GR ENG SIS TELECOMUN

GR ENG SIST AEROESP

GR ENG TELEMÀTICA

MU AEROSPACE S&T 09

MU DRONS

MU MASTEAM 2009

MU MASTEAM 2015

Cerca projectes

Calendari TFG de dipòsit i lectura

Documentació

Web UPC


 

Projecte llegit

Títol: Time to land prediction in Barcelona-El Prat airport based on machine learning classification models

Estudiants que han llegit aquest projecte:

Director: BARRADO MUXÍ, Cristina

Departament: DAC

Títol: Time to land prediction in Barcelona-El Prat airport based on machine learning classification models

Data inici oferta: 24-01-2020     Data finalització oferta: 24-09-2020


Estudis d'assignació del projecte:
    GR ENG AERONAVEGACIÓ
    GR ENG SIST AEROESP
Tipus: Coordinat    Títol: ATC time prediction based on machine learning models
 
Lloc de realització: EETAC
 
Paraules clau:
Machine learning , Airspace, ATC, Time to land, Prediction
 
Descripció del contingut i pla d'activitats:
This project aims at designing a tool based on machine learning models capable of predicting shortcuts given by ATC. To do so, several data will be obtained from Flightradar24.com thanks to the ADS-B antenna installed at the roof of the university which is located at about 7 km westbound from Barcelona airport.

There is not a specific time or criterion for ATC to give shortcuts. Nowadays, they depend on a “human factor” which is neither deterministic nor reliable. Machine learning will be the tool to extract a model capable of predicting those shortcuts in a consistent way. This model may be influenced by several factors such as: A/C type, weather, airline, speed, position, time of day, etc. It will also be fundamental to adjust the key parameters of the different classification models by training them. The used technique will consist on splitting the dataset into two parts, the first one as previously stated will be used to deduct the patterns and the second one will be used to check if the model is working as expected. The programming language to be used is Python since it has different libraries available such as SciKit-Learn or TensorFlow.

The benefits of giving shortcuts are: the reduction of the flight time thus reducing CO2 emissions, the possibility of compensating delays and the possibility of increasing the demand among others. On the other hand, the main drawbacks are that situational awareness is lost and although it may seem paradoxical if shortcuts are not predicted, too many planes may be entering a certain sector during a dense traffic hour (compared with EURCONTROL’s daily prediction) and it may be impossible for controllers to sustain the demand, converting in that way a benefit into a problem.
 
Overview (resum en anglès):
This document contains a method to assess the landing time in Barcelona-El Prat airport using machine learning classification models. The goal of this project is not to predict a continuous variable but if the time to land falls inside one of the following categories: advanced, delayed or planned. Additionally, two categories called very advanced and very delayed have been included containing flights that have had abnormal values on their time to land, either because of very fast approaches or very slow ones.
To obtain the data, an ADS-B antenna located on the aerospace and telecommunications engineering school of Castelldefels rooftop has been used. This antenna, captures the signals of all the arrivals into Barcelona which are later decoded thanks to an already existing program written in C#. By means of a custom program written in python, the most relevant characteristics of these flights are extracted and presented in a matrix format so that the different models can understand them.
All the data has been scaled and divided in training and test samples. The first ones are used to teach the different models and the second ones are used to measure their efficiency.
Six different models have been trained, four of them reached accuracy values over 60%, another one reached a value of 75% and another one could not go over 30%. The best three models, have been adjusted with two different techniques: random search and grid search. It has been verified that a random search is significantly better since the same results can be obtained but it requires much less time and computing resources. Also, different methods to enhance the results of the already adjusted models such as voting classifiers and boosters have been used.
Finally, oversampling techniques have been implemented to solve the five categories problem as the extreme cases are very underrepresented. Thanks to this technique, the accuracy of these two categories was improved but in turn the overall accuracy descended.
The performance of these models could be upgraded in the future if more flights or more characteristics of those flights were added.


Data de generació 26/01/2021