Improved Flight Time Predictions for Fuel Loading Decisions of Scheduled Flights with a Deep Learning Approach.

摘要: Under increasing economic and environmental pressure, airlines are constantly seeking new technologies optimizing flight operations to reduce fuel consumption. However, the current policy on loading, which has a significant impact aircraft weight, leaves room for improvement. Excess is loaded by dispatchers and(or) pilots ensure safety because of consumption uncertainties, primarily caused time cannot be predicted Flight Planning Systems (FPS). In this paper, we develop novel spatial weighted recurrent neural network model provide better predictions capturing air traffic information at national scale based multiple data sources, including Automatic Dependent Surveillance - Broadcast, Meteorological Airdrome Reports, airline records. model, adopt layers extract spatiotemporal correlations between features utilizing repetitive patterns interacting elements in aviation networks. A layer introduced learn origin-destination (OD) specific features, two-step training procedure integrate individual OD models into one network. This was trained tested using year historical from real operations. Results show that our can more accurate than FPS LASSO methods, especially flights with extreme delays. We also improved prediction, loading optimized 0.83% an example airline's fleet without depletion risk.