Research on real-time trajectory optimization methods for stratospheric airships based on deep learning

Abstract

Stratospheric airships are a type of large aircraft capable of operating for extended periods in the stratosphere. This paper focuses on real-time trajectory planning for stratospheric airships. It constructs an optimization path dataset based on the Gauss pseudospectral method and utilizes deep learning neural networks to solve the real-time path planning problem for stratospheric airships. The article first establishes a six-degree-of-freedom airship spatial motion model. It uses the Gauss pseudospectral method to transform the original optimization problem into a parameter optimization problem, which is then solved using sequential quadratic programming. During the ascent phase, based on the airship's speed, yaw angle, and pitch angle when transitioning from the troposphere to the stratosphere, a total of 26,901 optimized paths are generated using the Gauss pseudospectral method, and the influence of different initial states on the optimized paths is analyzed. During the level flight phase, 3960 optimized paths are generated based on different initial speeds and yaw angles, and an analysis of the impact of the initial yaw angle on the optimized paths is conducted. Finally, the dataset generated by the Gauss pseudospectral method is divided into training and testing sets. Long short-term memory (LSTM) networks and Transformer networks are employed to learn and generate optimized paths from the dataset. Comparison results show that the neural network model is highly consistent with the optimized paths obtained using the Gauss pseudospectral method. Furthermore, the path generation time is reduced from hundreds of seconds to seconds, leading to a significant improvement in generation time stability.