Introduction
Strategic Traffic Management Initiatives (TMIs), particularly Ground Delay Programs (GDPs), are essential tools for mitigating the operational costs associated with imbalances between air traffic demand and airport capacity. When an airport's capacity is compromised—often due to adverse weather—controllers must decide whether to hold aircraft on the ground at their origin or allow them to depart and potentially face expensive airborne holding patterns. While ground delays are safer and more cost-effective, planning them effectively requires accurate forecasts of future airport capacity.
In practice, airport capacity is highly uncertain. Traditional predictive models often struggle with distribution shifts or errors, leading to suboptimal GDP assignments. This research addresses these challenges by proposing a fully integrated framework that combines deep learning-based probabilistic forecasting with a Distributionally Robust Optimization (DRO) approach to the multi-airport ground holding program (MAGHP).
Methodology
Probabilistic Capacity Prediction
The first stage of the framework involves a deep learning-based prediction model designed to forecast arrival and departure capacity distributions. Using a Multilayer Perceptron (MLP), the model generates forecasts for a 12-hour horizon in 15-minute increments. Unlike deterministic models that provide a single point estimate, this probabilistic approach outputs the likelihood of various capacity levels, providing a richer input for the subsequent optimization stage.
Distributionally Robust Optimization (DR-MAGHP)
The core of the framework is the Distributionally Robust Multi-Airport Ground Holding Program (DR-MAGHP). Standard stochastic programming (SP-MAGHP) optimizes performance based on a fixed, assumed probability distribution. However, if the underlying distribution shifts—due to unmodeled weather phenomena or operational changes—the stochastic solution can become inefficient.
DR-MAGHP addresses this by optimizing for the worst-case distribution within a "neighborhood" of the predicted distribution. This neighborhood is defined as a Wasserstein ambiguity set, which uses the Wasserstein distance to measure the discrepancy between probability distributions. By considering a range of plausible distributions around the forecast, the model ensures resilience against prediction errors.
Computational Tractability
Managing flight paths across a network of 30 major airports presents a significant computational challenge. To maintain tractability, the researchers implemented a scenario reduction technique. This involved Wasserstein-based clustering to aggregate similar time periods and K-means clustering to simplify capacity distributions, transforming a potentially intractable problem into one solvable by standard optimization software.
Data and Experimental Design
The framework was evaluated using a comprehensive dataset from 2019, focusing on the FAA Core 30 airports. Data sources included:
- Throughput Data: Actual arrival and departure times from the Bureau of Transportation Statistics (BTS).
- Weather Data: Meteorological variables (e.g., visibility, wind speed, ceiling) from the High-Resolution Rapid Refresh (HRRR) database.
- Capacity Estimation: Since "true" capacity is not directly observed, the authors utilized a multi-criteria rule-based method to estimate capacity during "saturated" periods where demand exceeded throughput.
Results and Analysis
The performance of DR-MAGHP was compared against both stochastic (SP-MAGHP) and deterministic baselines.
Resilience to Distribution Shifts
The primary advantage of the DRO approach is its performance under uncertainty. In scenarios where actual airport capacity was 30% to 50% lower than predicted—a common occurrence during severe weather events—DR-MAGHP achieved a 9.95% to 15.6% improvement in total cost reduction compared to SP-MAGHP.
Comparison with Deterministic Models
When compared to deterministic "best guess" models, the robust framework showed even more dramatic gains, reducing operational costs by over 40% in adverse conditions. This highlights the significant risk associated with ignoring uncertainty in air traffic planning.
The Role of Strategic Conservatism
Analysis of the results indicates that DR-MAGHP tends to favor slightly more conservative ground delay assignments. While this increases initial ground delay costs, it prevents the much higher costs associated with airborne delays that occur when optimistic plans fail to account for capacity drops. This "safety bubble" created by the Wasserstein ambiguity set allows the system to remain robust even when the machine learning model is overconfident.
Conclusion and Future Work
This research demonstrates that integrating deep learning with distributionally robust optimization can significantly enhance the resilience of air traffic management systems. By accounting for "uncertainty about uncertainty," the DR-MAGHP framework provides a more reliable foundation for planning Ground Delay Programs.
Future research will focus on several key areas:
- Inter-airport Dependencies: Moving beyond the assumption of airport independence to model the "ripple effects" of delays across the network.
- Dynamic Robustness: Tailoring the size of the ambiguity set to specific airports based on their historical predictability.
- Arrival-Departure Correlation: Integrating the shared runway constraints that link arrival and departure capacities.
- Real-world Validation: Testing the framework against historical "disruption days" to further validate its performance in complex, real-world operational environments.
Summary updated