Drone delivery systems have proven to be a promising alternative for urban applications that require speed, operational flexibility, and adaptability to different demand patterns. In such scenarios, drones can contribute to reducing travel times, expanding service coverage, and supporting services that depend on faster deliveries. However, demand variability makes it difficult to adequately size the fleet, since periods of low demand can lead to underutilization of resources, while periods of higher intensity can cause delays and violations of Service Level Agreements (SLAs), especially those associated with average delivery time. This dissertation proposes Drones-DT, a Digital Twin architecture based on a Stochastic Petri Net model for the dynamic management of drone fleets in delivery systems. Drones-DT is continuously synchronized with a drone simulator, allowing the representation of the system's operational state and the performance of what-if predictive analyses to estimate performance metrics under different fleet configurations and demand conditions. Based on these estimates, an SLA-driven decision mechanism dynamically adjusts the number of drones in operation, seeking to balance meeting time requirements with the rational use of available resources. The results indicate that the proposed approach manages to maintain SLA compliance while reducing the average number of drones used when compared to static strategies. Therefore, Drones-DT demonstrates potential to support operational decisions in drone delivery systems, allowing the fleet to be adjusted according to environmental behavior and contributing to a more adaptive, controlled operation aligned with service needs.