Predicting Flight Departure Delay Durations Using Ensemble Learning: A Case Study of Soekarno-Hatta International Airport

Abstract

Flight delays at primary hubs like Soekarno-Hatta International Airport (CGK) can disrupt national connectivity and incur substantial operational costs. While existing research often relies on binary classification, tactical airport management requires precise temporal granularity in minutes to optimize resource allocation, such as gate and stand management. This study develops a robust duration prediction model using ensemble learning (XGBoost and Random Forest) integrated with a cost-sensitive learning strategy to address the severe skewness in delay duration distribution. The methodology incorporates advanced preprocessing, including Winsorizing to stabilize gradients and cyclical encoding to capture temporal continuity. Experimental results using 2024 operational data show that the optimized XGBoost model achieves superior performance with a Mean Absolute Error (MAE) of 6.39 minutes and an R² score of 0.70. Feature importance analysis identifies scheduled turnaround and ground infrastructure readiness as the primary determinants of delays, highlighting a significant "knock-on effect" where narrow transition windows fail to absorb inbound disruptions. These findings facilitate a transition from reactive reporting to proactive analytics, enabling the Airport Operation Control Center (AOCC) to optimize gate assignments and mitigate delay propagation.