Using Data-Driven Approach in 4D Trajectory Prediction: A Comparison of Common AI-Based Models

Abstract

Artificial intelligence (AI) is developing strongly and widely applied in many fields, including in the aviation. At the 13th Air Navigation Conference (AN-Conf/13-WP/232) organized by ICAO in Montreal from 9 to October 19 2018, participants discussed AI benefits and preparation for AI-enabled air traffic management (ATM). In this paper, a comparison was carried out to evaluate common AI-based models: Linear regression (LR), Random forest (RF) regression, Extremely gradient boosting (XGBoost) regression and deep neural network (DNN) models for predicting four-dimensional (4D) flight trajectory under weather uncertainties. The datasets used in this paper contain actual ADS-B historical trajectory data of flights from Ho Chi Minh to Ha Noi for 14 days (November 12–26 2021) and time-synchronized weather data along the waypoints of each flight. After comparing the training performance of LR, RF, XGBoost, DNN models, the best-fit DNN models were chosen for further improvement. By tuning their main hyperparameters, the training results are significantly improved in terms of training time and mean absolute errors.