[Paper Review] Passenger Flow Predictions at Sydney International Airport: A Data-Driven Queuing Approach

Harold Nikoue, Aude Marzuoli+3 more|arXiv (Cornell University)|Aug 20, 2015

Transportation Planning and Optimization7 references16 citations

TL;DR

This paper proposes a data-driven queuing model to predict passenger flow and delays at Sydney International Airport's immigration checkpoint using flight schedules, historical immigration timestamps, and staffing data. By integrating mixture models for passenger walk times and congestion-based staffing, the model accurately simulates queue lengths and wait times, with validation showing strong alignment to real-world data on key days, though underestimation occurs due to data gaps in desk availability.

ABSTRACT

Time spent in processing zones at an airport are an important part of the passenger's airport experience. It undercuts the time spent in the rest of the airport, and therefore the revenue that could be generated from shopping and dining. It can also result in passengers missing flights and connections, which has significant operational repercussions. Inadequate staffing levels are often to blame for large congestion at an airport. In this paper, we present a stochastic simulation that estimates the operational uncertainty in passenger processing at immigration. Congestion and delays are estimated on arrivals and departures based on scheduled flight departures and arrivals. We demonstrate the use of cellular tracking data in refining the model, and an approach to controlling congestion by adjusting staffing levels.

Motivation & Objective

• To model passenger arrival patterns and delays at Sydney International Airport's immigration checkpoint using real operational data.
• To address the challenge of variable passenger demand and staffing levels throughout the day, particularly during peak hours.
• To develop a simulation framework that integrates flight schedules, passenger counts, and service rates to predict queue dynamics.
• To validate the model against actual wait times and queue lengths derived from DWELL and DIMIA datasets.
• To improve airport operations by enabling proactive staffing adjustments based on predicted congestion.

Proposed method

• Utilizes Flight Information Display System (FIDS) data to obtain scheduled and estimated flight times for arrivals and departures.
• Integrates historical passenger timestamp data from the Australian Department of Immigration and Multicultural and Indigenous Affairs (DIMIA) to estimate service rates per immigration desk.
• Applies mixture models to generalize passenger walk times, removing dependency on specific gate locations.
• Implements a congestion-based staffing policy where servers are opened or closed based on queue length thresholds to maintain stable queue levels.
• Simulates daily passenger flows by combining flight occupancy distributions, arrival patterns, and dynamic staffing decisions.
• Validates predictions against actual wait times and queue lengths from the DWELL system, using 12 simulated days including July 25, July 26, and December 11, 2012.

Experimental results

Research questions

• RQ1 How accurately can a data-driven queuing model predict passenger wait times and queue lengths at Sydney Airport’s immigration checkpoint?
• RQ2 To what extent do variations in staffing levels and flight delays affect passenger congestion and service times?
• RQ3 How does the integration of FIDS and DIMIA data improve the realism and accuracy of passenger flow simulations?
• RQ4 What is the impact of data gaps—particularly in desk availability records—on model performance and prediction reliability?
• RQ5 Can a congestion-based staffing policy effectively control queue length and reduce wait times under real-world variability?

Key findings

• The model shows strong agreement with actual wait times on December 11, 2012, with slight underprediction in the morning and overprediction in the afternoon.
• On July 25, 2012, the model accurately captures the delay spike in the afternoon caused by delayed flights, though it overestimates delays around 3 PM due to incorrect desk count assumptions.
• Queue length predictions are generally consistent with observed trends, though the model overestimates queue size on July 26, likely due to inaccurate desk availability data.
• The simulation fails to predict queues when service rates exceed demand, indicating a limitation in modeling idle periods.
• Actual wait times derived from DWELL are likely a lower bound due to missing data, particularly between 1:30 PM and 2:00 PM on some days, which affects model validation.
• The model’s performance is significantly impacted by inconsistencies in manual recording of desk availability, highlighting data quality as a key challenge in real-world implementation.