Waving into the Future: Development of a Predictive Model for the Deployment of Airport Marshallers

Abstract

The objective of this study is to develop a predictive model that can accurately predict the marshaller task demand at an airport. Amsterdam Airport Schiphol is used as a case study for this research. Marshallers form a critical link between ground operations and airside operations, ensuring safe aircraft movements and a continuous operation. Findings indicate a misalignment between static staffing practices and dynamic operational demand. The aim is to build a data-driven model using historical data and influencing factors. ADS-B vehicle data, geofence polygons, aircraft arrival times, and engine testing records were used to determine the task demand. A process of spatial matching, task labelling, and segmentation was used to combine these datasets, after which the labelled segments were then combined into hourly counts. Resulting in an aggregated dataset with hourly task count that can be used in a machine learning model. LightGBM was implemented as a multi-output model that forecasts all task types jointly. The models were trained on nine months of data and performance was evaluated using the Mean Absolute Error, Root Mean Squared Error, Mean Error and the Coefficient of Determination. The models were validated each on their own forecasting horizons: 24 hours, 168 hours, and 2160 hours, and compared to two baselines: Seasonal naïve, and weekly hourly average. The results show that the predictive capability strongly differs between the task type. Docking is the only task that can be forecasted reliably, it follows daily patterns and has a clear link with aircraft arrivals. Docking shows stable performances over all horizons with RMSE values between 1.66 and 1.75 and MAE values between 1.26 and 1.35. The model outperformed the baselines on all prediction horizons and on all evaluation metrics. The other tasks did not show any valuable forecasting, with R2 values close to zero or negative. This indicates that these tasks are irregular or occur in low volumes. Making them not suitable to provide reliable hourly forecasts with the current data. Future work could assess if additional operational factors improve the predictive structure of the other three tasks. A longer dataset may also reveal patterns that are not visible in the current nine month data. Other resolutions such as 15-minute or 30-minute time intervals, or shift intervals may be relevant.