Interpretable machine learning for prediction of aircraft turnaround times

Halmesaari, Eppu

Interpretable machine learning for prediction of aircraft turnaround times

Files

master_Halmesaari_Eppu_2020.pdf (3.7 MB)

Perustieteiden korkeakoulu | Master's thesis

Unless otherwise stated, all rights belong to the author. You may download, display and print this publication for Your own personal use. Commercial use is prohibited.

Authors

Halmesaari, Eppu

Date

2020-12-15

Major/Subject

Machine Learning, Data Science and Artificial Intelligence

Mcode

SCI3070

Degree programme

Master’s Programme in Information Networks

Language

en

Pages

100+6

Abstract

Optimized aircraft turnaround processes are crucial for airlines to maximize the utilization of single aircrafts, but overly optimized scheduled turnaround durations may compromise schedule adherence. Predictions about the duration and information related to the drivers of the turnaround duration provide operatively valuable information as they enable reactive measures to be taken to mitigate possible delays. In addition, by revealing general patterns and potential bottlenecks in a turnaround process, this information can be used to improve individual subprocesses in the longer term. This thesis presents a proof-of-concept of an operative model which creates interpretable predictions of single turnarounds by showing the contributions of different factors on the predicted turnaround duration. Main elements of the model are the tree-based machine learning algorithm XGBoost and interpretability framework SHAP. Explanations are provided both on the local and global level which means that they can be used to interpret individual turnarounds and to deduce general patterns between variables in the dataset. Dataset of this study consists of variables such as passenger count and the amount of cargo which are used to describe different subprocesses. However, explicit data about the durations of subprocesses is not included in the dataset. From a scientific perspective the combination of such dataset and the methods mentioned, this thesis provides a novel explainable aggregated approach on the turnaround duration prediction by viewing the turnaround as a black box and by inferring the effects of different variables on the total duration. Results of this study imply that an explainable aggregated approach is capable of explaining the turnaround duration to a certain extent. The proof-of-concept model enables the deduction of the main patterns in the data and also provides useful information related to individual turnarounds. However, additional features are needed to be able to provide a more thorough understanding of the turnaround process in this kind of approach.

Optimoitu lentokoneiden kääntöaika mahdollistaa yksittäisten lentokoneiden suuremman käyttöasteen, mutta liialti optimoitu kääntöaika voi vaarantaa lentokoneiden aikataulussa pysymisen. Näin ollen ennusteet yksittäisten kääntöjen kestoista sekä tieto kääntöprosessin merkittävimmistä ajureista mahdollistavat myöhästymistä ehkäisevät toimenpiteet. Lisäksi ymmärrys yleisistä riippuvuussuhteista ja pullonkauloista käännön alaprosessien välillä mahdollistaa prosessien kehittämisen pidemmällä aikavälillä. Tämä työ esittelee operatiivisen työkalun mallin, joka luo selittäviä ennusteita yksittäisistä lentokoneen kääntöajoista tuomalla esiin yksittäisten selittävien tekijöiden vaikutuksen lopullisessa ennusteessa. Mallin pääosat ovat koneoppimisen algoritmi XGBoost sekä selittävyyttä tuova menetelmä SHAP. Selittävyyttä kääntöprosessin kestoon tuodaan sekä yksittäisen käännön tasolla että koko käytössä olevan datan tasolla. Näin ollen on mahdollista tutkia myös yleisiä riippuvuussuhteita muuttujien välillä. Käytettävissä oleva data sisältää muuttujia, joiden välityksellä pyritään kuvailemaan yksittäisiä käännön alaprosesseja, mutta tarkkaa tietoa yksittäisten alaprosessien kestosta ei ole saatavilla. Tieteellisestä näkökulmasta tämän tyyppisten menetelmien ja datan yhdistäminen luo uniikin lähestymistavan kyseiseen ongelmaan. Lähestymistapaa voidaan kutsua selittäväksi aggregaattimalliksi, jossa itse kääntöprosessin keston määräytymistä kohdellaan mustana laatikkona ja jonka käyttäytymistä pyritään selittämään erilaisilla muuttujilla. Tulosten perusteella selittävä aggregaattimalli kykenee antamaan selittäviä ennusteita käännön kestosta. Malli tuo esille hyödyllistä tietoa kuvailemalla yksittäisen käännön merkittävimpiä tekijöitä ja lisäksi se tuo esille yhteyksiä muuttujien välillä. Tulosten perusteella voidaan myös todeta, että tässä työssä käytettävissä olevat muuttujat kykenevät selittämään ilmiötä vain rajallisesti ja esimerkiksi paremman ennustustarkkuuden saavuttaminen vaatii uusia muuttujia ja dataa.

Supervisor

Luoma, Jukka

Thesis advisor

Tiilikainen, Mika

Keywords

turnaround, aggregated, predictive, interpretability, machine, learning

Permanent link to this item

https://urn.fi/URN:NBN:fi:aalto-2020122056351

Collections

[dipl] Perustieteiden korkeakoulu / SCI

Show all metadata

Interpretable machine learning for prediction of aircraft turnaround times

Files

URL

Journal Title

Journal ISSN

Volume Title

Authors

Date

Department

Major/Subject

Mcode

Degree programme

Language

Pages

Series

Abstract

Description

Supervisor

Thesis advisor

Keywords

Other note

Citation

Permanent link to this item

Collections