Machine Learning Air Traffic Delays Prediction Models

This repository includes the source code of Machine learning models for predicting air traffic delays at different levels, which are flight-level, airport-level, and network-level. The repository contains the following models:

1. Random Forest model: Predicts arrival delays for individual flights
2. LSTM model: Predicts aggregated arrival and departure delays for a single airport
3. DST-GAT model: Predicts aggregated arrival and departure delays of all airports a network

Models

Individual flight prediction

A Random Forest regression model was used to obtain flights' arrival delays at an airport. Features such as airline, planned arrival time, and airport capacity were used as input to predict the target variable (arrival delay).

The code for this model can be found in the notebook: flight_arr_delay_rf.ipynb.

Single airport prediction

Single airport delay prediction focuses on the arrival and departure delays of any given airport using a long-short term memory recurrent network (LSTM) model.

The code for this model can be found in the notebook: network_delay_gnn.ipynb.

Graph Neural Network

The dynamic spatial-temporal graph attention network (DST-GAT) is a model which is used to predict the delays for a network of airports. This is done based on the connection between airports and the history of every airport. For every airport, the departure and arrival delays are predicted for several lookahead time steps.

The code for this model can be found in the notebook: network_delay_gnn.ipynb.

Python library dependencies

The following libraries are needed to run the code in this repository:

• numpy
• scipy
• pandas
• scikit_learn
• keras
• spektral
• tensorflow

Configuring TensorFlow GPU functionality (Optional)

Training neural networks and in particular, Graph Neural Networks tends to be quite slow. We therefore also recommend configuring TensorFlow to work with a graphics card (CUDA compatible NVIDIA GPU required). To do this some additional dependencies should be installed (see Tensorflow GPU guide). Importantly, all the versions should be compatible. Figuring out which versions to install can be tricky. At the time of submission, the following versions were used and should be compatible:

• Tensorflow 2.7.0
• CUDA Toolkit 11.2
• cudnn 8.1.0

Details on where to find and how to install the dependencies can be found here:

• Tensorflow GPU guide
• cudnn install guide

Data

Currently, the tools use 2 main sets of data:

• EUROCONTROL R&D data: not provided with this software. You need to request a license and download the data on your own.
• Airport coordinates: provided in tools.constants.py for European top 50 airports.

Many of the functions in tools/extract.py have generate, write, and read capabilities, which means they generate the full data on the first cold run and return the stored filtered data from the data on subsequent runs.

Your file tree should look something like this:

[repostiont root]
├───data
│   ├───2018
│   │   ├───201803
│   │   ├───201806
│   │   ├───201809
│   │   └───201812
│   ├───2019
│   │   ├───...
│
├───[other files]
.   .
.
.
.

Contributors

This software was created and extended upon a TU Delft's Capstone Project (TI3150TU) for the minor Engineering with AI. The following people contributed to the code:

• Constantinos Aristodemou @ConstantinosAr
• Vlad Buzetelu @vladbuzetelu
• Tristan Dijkstra @IrTrez
• Theodor Falat @theofalat
• Tim Hogenelst @TimGioHog
• Niels Prins @Niels-Prins
• Benjamin Slijper @BenjaminSlijper
• Junzi Sun @junzis (project tutor)