Runway utilisation is a function of actual yearly runway throughput and annual capacity. The aim of the analysis in this project is to find data driven prediction models based on the features and relevant scenarios that might impact runway utilisation. The Gradient Boosting machine learning method will be assessed on their forecast performance and computational time for predicting the procedural and non-procedural runway exit to be utilised after the landing rollout. The Gradient Boosting method obtained an accuracy of 79% and was used to observe key related precursors of unique data patterns. Tests were conducted using runway and final approach data consisting of 54,679 arrival flights at Vienna airport.

Accidents on the runway triggered the development and implementation of mitigation strategies. Therefore, the airline industry is moving toward proactive risk management, which aims to identify and predict risk precursors and to mitigate risks before accidents occur. For certain predictions machine learning techniques can be used. Although many studies have explored and applied novel machine learning techniques on different radar and A-SMGCS data, the identification and prediction of abnormal runway occupancy times and the observation of related precursors are not well developed. In our previous papers, three existing methods were introduced, lasso, multi-layer perception, and neural networks, to predict the taxi-out time on the taxiway and the time to fly and true airspeed profile on final approach. This paper presents a new machine learning method where the existing machine learning techniques are combined for predicting the abnormal runway occupancy times of unique radar data patterns. Additionally the regression tree method is used in this study to observe the key related precursors extracted from the top 10 features. Compared with existing methods, the new method no longer requires predefined criteria or domain knowledge. Tests were conducted using final approach radar data and A-SMGCS runway data consisting of 78,321 flights at Paris Charles de Gaulle airport and were benchmarked against 500,000 flights at Vienna airport.

Accurate taxi-out time predictions are a valuable asset in enabling efficient runway scheduling in real-time operationsso as to reduce taxi-out times and fuel consumption onthe airport surface. This paper will focus on how the neural networks, regression tree, reinforcement learning, and multilayer perceptron methods can be used for predicting taxi-out time. These four methods are assessed based on their performance indicators, applied on Charles de Gaulle operational taxi data and benchmarked against real-life taxi-out time profiles. The root-mean-squared error metric is chosen as the most important performance indicator, which gives, for the applied regression tree method, on any given day, an average error of 1.6 min. The regression tree turns out to be the most efficient method, which is then subsequently applied in a case study for predicting the taxi-out time and finding the key-related precursors extracted from the top 10 features.