Time Series Clustering of Weather Observations in Predicting Climb Phase of Aircraft Trajectories

Reliable trajectory prediction is paramount in Air Traffic Management (ATM) as it can increase safety, capacity, and efficiency, and lead to commensurate fuel savings and emission reductions. Inherent inaccuracies in forecasting winds and temperatures often result in large prediction errors when a deterministic approach is used. A stochastic approach can address the trajectory prediction problem by taking environmental uncertainties into account and training a model using historical trajectory data along with weather observations. With this approach, weather observations are assumed to be realizations of hidden aircraft positions and the transitions between the hidden segments follow a Markov model. However, this approach requires input observations, which are unknown, although the weather parameters overall are known for the pertinent airspace. We address this problem by performing time series clustering on the current weather observations for the relevant sections of the airspace. In this paper, we present a novel time series clustering algorithm that generates an optimal sequence of weather observations used for accurate trajectory prediction in the climb phase of the flight. Our experiments use a real trajectory dataset with pertinent weather observations and demonstrate the effectiveness of our algorithm over time series clustering with a k-Nearest Neighbors (k-NN) algorithm that uses Dynamic Time Warping (DTW) Euclidean distance.