Data-driven Methods to Capture Network-wide Traffic Uncertainties for Better Flight Time Prediction and Fuel Loading Decision

Abstract

Fuel loading decision is crucial for airline operations, as it directly impacts flight safety, aircraft emissions, and operational cost. In the current practice of fuel loading, a large amount of unused fuel may remain in the tank after landing, which results in economic and environmental waste. Dispatchers and pilots tend to load excessive fuel to ensure flight safety and to compensate for the uncertainties of flight time estimation by the current flight planning system (FPS). Trip fuel is an essential fuel loading category to cover fuel consumption of the entire flight journey, but its quantity estimation is sensitive to the dynamic traffic conditions, while current FPS fails to capture significant flight uncertainties from the air traffic network. In current policy, a part of fuel loading categories, including contingency fuel, discretionary fuel, and some extra fuel, is loaded to overcome unforeseen consumption uncertainties due to this limitation of the current flight planning model. This part of fuel has the potential to improve fuel loading decisions by providing airlines, dispatchers, and pilots with more accurate network delay estimations and individual flight time predictions. Although previous studies have identified room for optimization in reducing unnecessary fuel carrying and much of existing literature introduces novel methods to better predict flight delays, few works directly explore how to better embed network traffic status estimations into flight time predictions and further improve the operational flexibility and fuel loading strategies.

To better understand and capture the uncertainties in air traffic network and further support fuel loading decisions at the flight planning stage, three works are done: First is to explore the improvement opportunity to incorporate the nationwide air traffic information into individual flight predictions. In this work, we proposed a novel flight time prediction model based on a recurrent-based deep learning neural network to extract network-based air traffic information from historical data resources and predict en-route flight time for individual flights. Results show that the proposed model can provide more accurate flight time predictions than baseline methods, especially for flights with extreme delays. With the improved flight time prediction, fuel loading can be optimized to reduce fuel consumption by 0.026%- 2.661% for the example airline's fleet without increasing the fuel depletion risk.

In the second work, rather than one-point estimation at a constant prediction horizon, probabilistic estimations of flight time with multiple prediction horizons {are more proper to support dispatchers' fuel loading decision-making and allow the flexibility of payloads adjustment in airline operation.} In this work, we propose a novel framework to integrally make multi-horizon probabilistic predictions for a scheduled flight while considering the previous predictions' influence on following predictions with changing the transmission state. Results show that the proposed model outperforms other baselines, and ablation studies validate the efficacy of introduced modules.

The third work is to identify and characterize the network spatiotemporal correlations within airports. We propose a deep-learning-based model framework stacking graph attention layers and recurrent neural networks for nationwide airport throughput prediction. Two kinds of air traffic graph modeling methods with the airport as nodes and OD-pair as nodes are also investigated. Results show that in our proposed layer-stacking structured neural network, the temporal correlations are more significant than spatial ones in airport traffic predictions. Besides, the learned weights of our model can demonstrate the interacted relations changed over time among airports, which can help dispatchers and pilots better understand network traffic patterns and adjust fuel loadings accordingly.

Aiming to bridge the research gap to improve airlines' fuel loading decisions at the flight planning stage through better capturing network-wide air traffic uncertainties, the thesis investigates the opportunity based on data-driven methods to provide more accurate network delay estimations and individual flight time predictions. The analysis and implications in the thesis provide insights to aviation stakeholders and help airline dispatchers and pilots to better understand the air traffic dynamics, as well as further go for fuel loading strategies and other applications to make a more sustainable and intelligent air transportation system.

Date of Award	13 Feb 2023
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Lishuai LI (Supervisor)