[Paper Review] Predicting Aircraft Trajectories: A Deep Generative Convolutional Recurrent Neural Networks Approach

Yulin Liu, Mark Hansen|arXiv (Cornell University)|Dec 30, 2018

Air Traffic Management and OptimizationEngineering51 citations

TL;DR

The paper presents a deep generative CNN-RNN model that predicts 4D aircraft trajectories by modeling track points as conditional Gaussian mixtures, using weather-derived image features and an LSTM encoder–MDL decoder, with beam search and Kalman-based smoothing during inference.

ABSTRACT

Reliable 4D aircraft trajectory prediction, whether in a real-time setting or for analysis of counterfactuals, is important to the efficiency of the aviation system. Toward this end, we first propose a highly generalizable efficient tree-based matching algorithm to construct image-like feature maps from high-fidelity meteorological datasets - wind, temperature and convective weather. We then model the track points on trajectories as conditional Gaussian mixtures with parameters to be learned from our proposed deep generative model, which is an end-to-end convolutional recurrent neural network that consists of a long short-term memory (LSTM) encoder network and a mixture density LSTM decoder network. The encoder network embeds last-filed flight plan information into fixed-size hidden state variables and feeds the decoder network, which further learns the spatiotemporal correlations from the historical flight tracks and outputs the parameters of Gaussian mixtures. Convolutional layers are integrated into the pipeline to learn representations from the high-dimension weather features. During the inference process, beam search, adaptive Kalman filter, and Rauch-Tung-Striebel smoother algorithms are used to prune the variance of generated trajectories.

Motivation & Objective

• Develop a generalizable pipeline to convert high-fidelity meteorological data into image-like features for trajectory prediction.
• Model trajectory points as conditional Gaussian mixtures learned by an end-to-end neural network.
• Leverage an LSTM encoder to ingest flight plan information and a mixture-density LSTM decoder to capture spatiotemporal correlations.
• Incorporate convolutional layers to extract representations from high-dimensional weather features.
• Improve inference with beam search, adaptive Kalman filtering, and Rauch-Tung-Striebel smoothing to refine trajectory variance.

Proposed method

• Construct image-like feature maps from wind, temperature, and convective weather using a tree-based matching algorithm.
• Use an end-to-end convolutional recurrent network with an LSTM encoder and a mixture density LSTM decoder.
• The decoder outputs parameters of Gaussian mixtures for trajectory points.
• Encoder embeds last-filed flight plan information into fixed-size hidden states.
• Convolutional layers learn representations from high-dimensional weather features.
• During inference, apply beam search, adaptive Kalman filter, and RTS smoother to prune trajectory variance.

Experimental results

Research questions

• RQ1 Can image-like meteorological feature maps be effectively used for 4D trajectory prediction?
• RQ2 Can a deep generative model learn reliable Gaussian mixture outputs for aircraft tracks conditioned on flight plans and weather?
• RQ3 Does integrating convolutional weather representations with an LSTM-based decoder improve trajectory accuracy and uncertainty estimation?
• RQ4 How do inference-time techniques (beam search, Kalman filter, RTS smoothing) affect trajectory quality and variance reduction?

Key findings

• Proposes a generalizable pipeline to transform weather data into image-like features for trajectory modeling.
• Models trajectory points as conditional Gaussian mixtures whose parameters are learned by a deep generative model.
• Demonstrates end-to-end learning with an LSTM encoder and a mixture-density LSTM decoder.
• Incorporates convolutional layers to capture high-dimensional weather feature representations.
• Employs beam search, adaptive Kalman filtering, and RTS smoothing during inference to prune variance.