Abstract:Flight trajectory prediction is the basis of air traffic management and simulation. Aiming at the problems of insufficient accuracy and instability of existing prediction methods, a trajectory prediction algorithm based on historical radar data and machine learning was proposed in this paper. The positional offsets between adjacent update moments were regarded as the model object and Gaussian mixture model (GMM) was applied to model the distribution of data. Then the historical flight trajectory was used to optimize the model parameters by expectation maximization (EM) algorithm. The probabilistic statistical model can precisely describe the randomicity of flight operation by learning the moving pattern of the same flight with collected trajectory. Based on the learned model, the sequential Monte Carlo (SMC) approach was used to sample the relative position offsets sequence at each update time during flight operation. For a given flight, the position and elevation of the runway at the departure airport are the initialization of trajectory. Combined with sequence of positional offsets, the position of the flight was predicted, which further generates a complete flight trajectory. The model parameters were updated based on the collected trajectories in the database when the prediction errors exceed the preset threshold, or over a regular period of time. The proposed method has been applied to a real air traffic flow management system. The results showed that the prediction results are more accurate than that of kinematic model and regression model. Moreover, during the climb and descent phase of the flight, the prediction results are more stable than that of the existing methods, which is a fatal bottleneck of these methods.