Williams_Capacity_AMSARAM_20250116_final.pptx
Neural Network Prediction of Airport Arrival and Departure Capacity
John K. Williams, Chris Dorsey, Kate Collins, Scott McCall, Serhiy Likhanov, and Emily Schofield*
The Weather Company
Introduction
This poster describes the development of predictive models for Airport Acceptance Rate (AAR) and Airport Departure Rate (ADR) capacity: hourly limits on how many aircraft can arrive or depart an airport, respectively. Incorporated into a real-time decision support system, these forecasts allow airline dispatchers to proactively mitigate potential disruptions.
Data
Predictand: Inferred AAR & ADR
In contrast to previous studies based on historical "called" AAR and ADR capacity from the National Traffic Management Log, we objectively identify cases of near-capacity or degraded capacity operations from historical arrival and departure data from 99 airports in the US, Canada and Mexico for training and evaluation. Degraded capacity.
| Ceiling | QPF Snow |
|---|---|
| Cloud cover | Relative Humidity |
| cos(DayOfYear) | Scattered CBH |
| DayOfWeek | Temperature |
| Dewpoint | UV Index |
| Dewpoint Depression | Visibility |
| iconCode | Wind Dir Delta |
| PrecipProb | Wind Direction |
| Pressure MSL | Wind Gust |
| QPF | Wind Speed |
| QPF Ice | Wx Severity |
Predictors: TWC Forecast on Demand
We utilize a historical archive of FoD point forecasts at airport locations. FoD is TWC’s automated, on demand consensus forecast with a Human Over the Loop workflow. Temporal, weather forecast and derived predictors used in the study are shown in the table below.
Approach
We used the MATLAB Deep Learning toolbox to train a neural network model, selecting 3 hidden layers (128, 64 and 32 nodes) based on hyperparameter tuning. The model is exported in ONNX for real-time operational implementation.
The model was trained on forecast lead times of 1, 2 and 3 hours, but applied to all lead times.
Data Analysis
Maverick Beta Display
TWC’s Terminal Airspace Convection Risk product comprises 7-hour forecasts of arrival and departure gate impacts at 41 global airports.
Evaluation
RMSE and MAE units are % of airport capacity. Max CSI and AUC are for discriminating degradation of airport capacity below 85%.
| Airport | RMSE | MAE | Max CSI | AUC |
|---|---|---|---|---|
| KATL | 28% | 13% | 0.24 | 0.82 |
| KORD | 28% | 12% | 0.20 | 0.82 |
| KDFW | 28% | 12% | 0.32 | 0.82 |
| KDEN | 30% | 12% | 0.24 | 0.80 |
| KCLT | 31% | 14% | 0.34 | 0.87 |
| KLAX | 15% | 3% | 0.06 | 0.80 |
| KLAS | 29% | 12% | 0.22 | 0.76 |
| KJFK | 27% | 12% | 0.20 | 0.80 |
| KMIA | 27% | 11% | 0.26 | 0.82 |
Contact: john.williams@weather.com
Summary
We created neural network forecasts of hourly AAR and ADR based on observed airport arrival and departure counts collected between August 2022 and December 2024. Evaluation on an independent test dataset shows that ADR predictions generally score substantially better than AAR, likely because arriving flights are more sensitive to enroute weather not represented in the point location forecasts used as predictors. Future work will incorporate (1) enroute weather predictors from TWC’s TrACR product, (2) runway configuration predictions, and (3) National Airspace System parameters reflecting non-local and non-weather factors.