Airport surface dataset for machine learning based taxi-out time prediction

Sammanfattning

Airport surface operation is an important part of air traffic management. It is acritical contributor of overall flight efficiency. Accurate prediction of taxi-out timefor aircraft can help reduce delays and allow for more robust scheduling. This thesisuses airport surface movement data to predict taxi-out time using machine learning.By using radar and sensor data collected at one major airport, the goal is to createa dataset suitable for machine learning.To achieve this, a comprehensive preprocessing pipeline was developed to harmonizeairport data from A-SMGCS surveillance systems and ASTERIX-based systems.Key features influencing taxi times, such as traffic congestion, aircraft characteristics,airport infrastructure, temporal patterns, and weather conditions, wereidentified through a literature review. This thesis reconstructs the identified featuresusing the available airport data. Additionally, feature selection and multicollinearityanalysis were performed using correlation matrices, variance inflation factor (VIF),and Cramer’s V statistics to ensure the robustness of the predictive model.Finally, in order to evaluate the dataset, two supervised machine learning modelswere used. The models implemented were Multiple linear regression and Randomforest. Random forest outperformed linear regression, showcasing its ability to capturenonlinear and complex patterns in the dataset. For a final robust evaluationof the dataset, k-fold cross validation was used on Random forest. The results werethen interpreted using SHAP. SHAP identified features pertaining to the airportgeometry and congestion-related features as the features with the most impact onthe taxi-out time prediction.This study demonstrates the feasibility and limitations of creating an airport surfacedataset for machine learning for predicting taxi-out times. It highlights the importanceof the data preprocessing steps, as well as the feature engineering, in airportsurface-related predictive modeling.