FlightBERT++: A Non-autoregressive Multi-Horizon Flight Trajectory Prediction Framework

Introduction

This repository provides source codes of the proposed flight trajectory prediction framework, called FlightBERT++, and example samples for the paper FlightBERT++: A Non-autoregressive Multi-Horizon Flight Trajectory Prediction Framework. This work is proposed to i) forecast multi-horizon flight trajectories directly in a non-autoregressive way, and ii) improve the limitation of the binary encoding (BE) representation in our previous work FlightBERT.

Repository Structure

FlightBERT++
│  dataloader.py (Load trajectory data from ./data)
│  LICENSE (LICENSE file)
│  model.py (The neural architecture corresponding to the FlightBERT++ framework)
│  README.md (The current README file)
│  run.py (The main file for the model training and testing)
│  utils.py (Tools for the project)
├─data
│  │ README.md (README file for the dataset.)
│  │ example_data.txt (Example data file)
│  ├─dev (Archive for the validation data)
│  ├─test (Archive for the test data)
│  └─train (Archive for the training data)
└─pics

Package Requirements

• Python == 3.7.1
• torch == 1.9.0 + cu110
• numpy == 1.18.5
• matplotlib == 3.2.1

System Requirements

• Ubuntu 16.04 operating system
• Intel(R) Core(TM) i7-7820X@3.6GHz CPU
• 128G of memory
• 8TB of hard disks
• 8 $\times$ NVIDIA(R) GeForce RTX(TM) 2080 Ti 11G GPUs.

Instructions

Installation

Clone this repository

git clone https://github.com/gdy-scu/FlightBERT_PP.git

Create proper software and hardware environment

You are recommended to create a virtual environment with the package requirements mentioned above and conduct the training and test on the suggested system configurations.

Training and Testing

The training and testing are both packaged into the script of run.py for the FlightBERT++ framework with different arguments in config.json.

The main arguments in config.json are described below:

learning_rate: Float. The learning rate of the Adam optimizer. default=0.0001

period: Integer. The sampling period for dataloader. default=5

batch_size: Integer. The number of samples in a single training batch. default=2048

epoch: Integer. The maximum epoch for the training process. default=20

train_data: String. The path for the training set. default='./data/train/'

dev_data: String. The path for the validation set. default='./data/dev/'

test_data: String. The path for the test set. default='./data/test/'

saving_dir: Integer. The save path of the models and log file during the training/testing process. default='./check_points/'

n_en_layer: Integer. The layer number of the Transformer block in the encoder. default=4

n_de_layer: Integer. The layer number of the Transformer block in the decoder. default=4

horizon: Integer. The prediction horizons of the flight trajectory prediction task. default=15

is_training: Bool. Used to specify the running mode, true for training and false for testing. default=true

model_path: String. The checkpoint model path for the traning or testing. default=''

To train the FlightBERT++ framework, use the following command.

python run.py --config ./config.json

Test

To test the model, set is_training to false and set the model_path to the specific test model (config.json), and run the following command.

python run.py --config ./config.json

Dataset

In this repository, the example samples /data/example_data.txt are provided to facilitate quick start. The guidance about the example data can be found in /data/README.

Citation

Guo, D., Zhang, Z., Yan, Z., Zhang, J., & Lin, Y. (2024). FlightBERT++: A Non-autoregressive Multi-Horizon Flight Trajectory Prediction Framework. Proceedings of the AAAI Conference on Artificial Intelligence, 38(1), 127-134. https://doi.org/10.1609/aaai.v38i1.27763

Contact

Dongyue Guo (dongyueguo@stu.scu.edu.cn)