Modeling pre-tactical Air Traffic Flow Management decision processes with Maschine Learning/Deep Learning approaches

Kurzfassung

In the context of this master thesis,a structured analysis of a comprehensive set of flight plan data is carried out with the help of advanced analytical methods from the fields of Machine Learning and Deep Learning in order to draw conclusions about ATFM-decision processes in the pre-tactical phase regarding operational influence parameter. For this purpose, operational flight plan data for three months, June, August, and November for year 2016 was taken from EUROCONTROL RnD and Sabre Airport Intelligence Data database, and was processed and parameterized and fed into ML/DL model to predict the flight planned trajectories. The entire dataset used in thesis contained flight specific datapoints for 1,558,791 number of flights . The datasets where merged in a pre-processing step in order to use them as an input for the Machine Learning models. Before the merging process, the Sabre Airport Intelligence data first had to be translated from IATA to ICAO Codes After, translation the data was merged with EUROCONTROL RnD. During the merge, data loss was observed since data from Sabre Airport Intelligence Data did not contain the information on domestic flights in European region. Semi-Supervised Learning approach was taken, which is the combination of both Unsupervised Learning and Supervised Learning. The output labels used in the training and validation process are determined by the Unsupervised Learning technique called clustering. Hence, Unuspervised Learning was employed to generate the output labels by using a technique called clustering.In order to handle the large amount of data, an automated clustering technique called DBSCAN was selected. Before feeding into DBSCAN dataset was split randomly into training set which had 80% and test set which 20% of the dataset respectively. Clustering was performed on the training dataset, and the test dataset was transformed to the training set correspondingly. The DBSCAN algorithm determined the clusters labels for each O-D pairs by taking the lateral area between two trajectories and normalizing over the haversine distance between the corresponding OD pair as distance matrix value. A total of 21,199 clusters were formed for 16,304 unique OD pairs. In order to validate the performance and quality of the clusters created, clusters obtained for a representative OD pairs was plotted on the map. It was observed, in general the clustering algorithm produced decent results. While determining the clusters for log haul flights, in spite of having many branches it was clustered as single cluster. This was because of the lateral area for the long flights was bigger. On further investigation, plotting scatter plot by taking an average distance matrix for each OD pair against the average distance matrix to its corresponding cluster.58 5. Summary The plot confirmed again that long haul flight was below the curve fit over the data points. The reasons for the data points below the curve was that the average lateral area between those trajectories were more than the average distance matrix. Once the cluster labels were created for each flights, the dataset was fed into the Machine Learning algorithm. Using Supervised Learning techniques prediction of trajectory was carried out. The presented dataset was imbalanced data with multiclass and multi-label classification problem. Ensemble Learning was selected as a Supervised Learning in order to predict planned flight routes was selected over the popular Neural Network. Ensembling Learning combines several individual models to obtain better generalization performance. Also, multi-class and multi-label problems don’t need any sampling of the data. Ensembling Learning like Random Forest, XG Boost and Light GBM algorithms were implemented. Then compare the performance among each other and also with the PREDICT Logic. However, due to the computational constraint, only Random Forest was trained. The predictions created by the Random Forest algorithm showed an average f1-score of 82% Similarly f1-score for PREDICT Logic with the original cluster label, which performed 5% was less than the Machine Learning model. In general, the Ml model performed well predicting the labels, the model is biased towards precision of the model with slightly less recall. In conclusion, even though the prediction from the Machine Learning model performed well, the model’s recall is low. This could be improved by doing hyperparameter tuning, like performing Randomized Search Cross-Validation or Grid Search CrossValidation.