A Top of Descent Prediction Model for Interaction-free Continuous Descent Operations

When air traffic controllers select the Top of Descent (TOD) location for flights without considering downstream traffic interactions, the descent process may get interrupted with level-offs to avoid conflicts. To support green aviation practices such as Continuous Descent Operations (CDO), this paper proposes a two-step learning model to predict TOD locations that lead to interaction-free descent trajectories to enable continuous descents. The first step involves identifying and learning from non-CDO flights whose descents were interrupted due to flight interactions. This process models the critical areas—primary zones containing interacting flights that may disrupt flight descents. In the second step, the model learns from CDO flights that have successfully maintained CDO notwithstanding the presence of potential interacting flights in the critical areas that cross paths or converge with the flight. A random forest-based model is trained to understand how the relationships between focal and potential interacting flights (e.g., relative altitudes, distances, convergence points) influence TOD decisions. TOD prediction results on the major arrival flow in Singapore flight information region (FIR), using the Air Traffic Management System (ATMS) data for November 2019, show that above 88% of the predicted TOD locations are within ±10nm of the actual TOD of the CDO flights on the test dataset, with a Mean Absolute Error of 5.14 nm. Moreover, testing the model on non-CDO flights with leveling-offs caused by flight interactions demonstrates that the prediction model can help avoid flight’s leveling-offs by recommending a later TOD, allowing interacting flights to pass before the descent begins.