Taxi-Out Time Prediction Model at Charles de Gaulle Airport

Aircraft taxi time prediction: Feature importance and their implications

“…The ensemble machine learning, ordinary least-squared and penalized approaches were testified and the results suggest that no algorithm outperforms others in all cases, and one should strike a balance between the prediction bias and variance. Herrema et al [27] focused on Neural Networks (NN), Regression Tree (RT), RL and multilayer perceptron (MLP) methods for the taxi time prediction at Charles de Gaulle Airport (CDG). The top 10 out of 42 features, e.g., unimpeded taxi time, congestion level, and number of departures in the last 20 minutes were chosen in the feature selection process, and RT turned out to be the most efficient method.…”

“…Two features related to aircraft turning angles that may affect the taxi speed are included in the prediction model as well. Besides, aircraft weight has been considered as a potential candidate factor for taxi time prediction [11,27]. In line with the aircraft wake vortex [39], we introduced feature aircraft weight to categorise the aircraft as small, medium and large.…”

“…Another drawback we have observed falls in the feature importance identification process. Although as mentioned, various prediction methodologies have been applied in this domain, only a few of existing research addressed the identification of feature importance [26,27]. Assuming there is no priority information for the features, Jordan et al [26] applied a sub-optimal subset selection method [28] to capture much of the variability in the data source, though with no guarantee of ensuring high-quality prediction accuracy.…”

“…Assuming there is no priority information for the features, Jordan et al [26] applied a sub-optimal subset selection method [28] to capture much of the variability in the data source, though with no guarantee of ensuring high-quality prediction accuracy. Herrema et al [27] defined two metrics to rank the candidate features, and selected the top ten of them as the most important features. Note a sufficient analysis of the selected features number was not provided.…”

Machine Learning Techniques for Taxi-out Time Prediction with a Macroscopic Network Topology

“…For the other machine learning models, algorithms of reinforcement learning, support vector machines, k-nearest neighbors, random forest, and neural networks have been attempted to predict the taxi-out time. The system state, surface spot, runway, gate, departure fix and weight class, are selected as the features of machine learning methods [9][10][11][12]. Then, the performances of machine learning algorithms are analyzed and compared based on the computational results from the training and test experiments.…”