Stochastic Strategic Demand–Capacity Balancing Using Model-Free Reinforcement Learning

Abstract

Strategic demand-capacity balancing is essential for mitigating inefficiencies in air traffic operations, particularly under dynamic and uncertain weather conditions. This work proposes a stochastic demand-capacity balancing approach based on reinforcement learning, in which an agent allocates ground delays to individual flights using traffic demand and weather forecasts. The method is evaluated using real-world data for São Paulo/Guarulhos International Airport (SBGR), with performance assessed against observed weather conditions. An additional case study illustrates the applicability of the approach to Urban Air Mobility (UAM) operations, showing that the same algorithm can support traffic flow management in both conventional and emerging air traffic environments. The results highlight the potential of reinforcement learning to support adaptive, uncertainty-aware demand-capacity balancing in future Air Traffic Management (ATM) systems.

Method in brief

Strategic layer

Maskable PPO assigns per-flight ground delays in 15-minute increments, capped at one hour and by the remaining planning horizon.

MDP formulation

The state captures the previous operational configuration, traffic queues, weather category and wind state; the action is a per-flight ground delay, and the reward links strategic delay decisions to their downstream tactical congestion consequences.

Weather uncertainty

Training combines empirical weather dynamics with forecast-error information derived from TAF and METAR records.

Tactical layer

Reallocated demand is evaluated against runway or vertiport capacity constraints, allowing the strategic policy to internalize tactical congestion effects.

Evaluation

Conventional ATM is evaluated at SBGR using real-world traffic and weather data; the same learning framework is also examined in a UAM case study for the São Paulo Metropolitan Region.