Predicting Fuel Consumption and Flight Delays for Low-Cost Airlines

Yuji Horiguchi, Yukino Baba, Hisashi Kashima, Masahito Suzuki, Hiroki Kayahara, Jun Maeno

Abstract

Low-cost airlines (LCAs) operate with minimal margins, making operational efficiency paramount. This study investigates the use of machine learning models—Random Forests, XGBoost, and Deep Neural Networks—to predict fuel consumption and flight delays using real-world data from Peach Aviation. Our findings reveal that machine learning models, particularly XGBoost, can reduce fuel prediction errors by approximately 39.3% compared to human dispatchers. Furthermore, these models maintain a lower under-estimation ratio, enhancing safety while optimizing costs. Delay prediction models also show promise, achieving an AUC of 0.75 for key routes.

Introduction

For low-cost airlines, fuel represents the single largest operating cost. Traditional methods for estimating fuel consumption often rely on empirical rules used by human dispatchers, which frequently lead to over-estimation. While over-estimation ensures safety, the added weight of "buffer" fuel paradoxically increases total fuel burn. Conversely, under-estimation poses a significant safety risk.

Additionally, LCAs rely on high aircraft utilization. The "delay domino effect" means a single late flight can disrupt an entire day's schedule. Predicting these delays in advance is crucial for mitigating operational impact through proactive scheduling and resource allocation.

Methodology

The researchers treated fuel consumption as a regression task and flight delays as a binary classification problem (delay $\ge 15$ minutes). Three architectures were evaluated: Random Forests (RF), XGBoost (XGB), and Deep Neural Networks (DL).

A critical component of the methodology was feature engineering. Temporal data (dates and times) were transformed using trigonometric functions to capture cyclical patterns. Categorical data, such as airport and airframe IDs, were processed using one-hot encoding. Models were tested across three time horizons: one day (D), one week (W), and five months (M) before departure.

Data Sources

The study utilized a proprietary dataset from Peach Aviation covering July 2012 to March 2015. This included:

  • Flight Data: 54,000 flights across 12 airports.
  • Passenger Data: Demographics for 9.9 million passengers.
  • Reservation Data: Historical booking curves to account for load factors.

Results and Discussion

Fuel Consumption Optimization

The XGBoost model (XGB-D) demonstrated superior performance, achieving a relative Root Mean Squared Error (RMSE) of 8.8%, compared to 14.5% for human estimates (EX-D). This represents a nearly 40% improvement in accuracy.

Relative RMSE of fuel consumption predictions across various airports, comparing machine learning models against human expert estimates (EX-D).

Safety and Under-estimation

A key concern in automating fuel estimation is the risk of under-estimation. Analysis of the "under-estimate ratio" showed that the XGBoost model remained highly conservative, often maintaining a ratio below 1% for most airports. Interestingly, the ML models often exhibited lower under-estimation ratios than human dispatchers at several major hubs, suggesting that data-driven approaches can be both more accurate and safer.

Under-estimation ratios for different airports, showing that XGBoost (XGB-D) maintains higher safety standards compared to the human baseline (EX-D).

Flight Delay Prediction

Delay prediction proved more challenging, particularly for longer time horizons. However, for one-day-ahead predictions, the models achieved an AUC of 0.75 for specific routes like Tokyo-Narita (NRT) and Nagasaki (NGS). Feature importance analysis revealed that "fuel on board" was a significant predictor of delays, likely acting as a proxy for expected adverse weather or complex routing.

Conclusion

This research highlights the transformative potential of machine learning in the aviation industry. By providing more accurate fuel estimates and reliable delay predictions, ML models can help LCAs reduce costs and improve service quality. Future improvements could involve integrating real-time meteorological data and flight recorder information to further refine prediction accuracy.

Summary updated

Abstract

Low-cost airlines (LCAs) represent a new category of airlines that provides low-fare flights. The rise and growth of LCAs has intensified the price competition among airlines, and LCAs require continuous efforts to reduce their operating costs to lower flight prices; however, LCA passengers still demand high-quality services. A common measure of airline service quality is on-time departure performance. Be- cause LCAs apply efficient aircraft utilization and the time between flights is likely to be small, additional effort is required to avoid flight delays and improve their service quality. In this paper, we apply state-of-the-art predictive modeling approaches to real airline datasets and investigate the feasibility of machine learning methods for cost reduction and service quality improvement in LCAs. We address two prediction problems: fuel consumption prediction and flight delay prediction. We train predictive models using flight and passenger information, and our experiment results show that our regression model predicts the amount of fuel consumption more accurately than flight dispatchers, and our binary classifier achieves an area under the ROC curve (AUC) of 0.75 for predicting a delay of a specific flight route.
BibTeX
@inproceedings{horiguchi2017predicting,
  author = {Yuji Horiguchi and Yukino Baba and Hisashi Kashima and Masahito Suzuki and Hiroki Kayahara and Jun Maeno},
  title = {{Predicting Fuel Consumption and Flight Delays for Low-Cost Airlines}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2017},
  doi = {10.1609/aaai.v31i1.19095},
}
Figure