CEEMDAN-STimesnet: A Deep Learning Model for Low Earth Orbit Satellite Trajectory Prediction

In the trajectory prediction of low Earth orbit (LEO) satellites, orbital height serves as a critical metric. We employ the CEEMDAN-STimes -Net deep learning model to forecast the orbital height of LEO satellites for the next 36 hours. By utilizing CEEMDAN, the orbital height is decomposed into several subsequences, effectively separating the trend components of the data. The model integrates spatial attention with TimesNet, innovatively transforming one-dimensional time series into multi-cycle two-dimensional tensors, ensuring the efficient extraction of historical information. The introduction of spatial attention enhances the model’s emphasis on salient spatial features within the two-dimensional channels. Experimental results demonstrate that compared to the Transformer and its variants, the model achieves a 57.55% reduction in Mean Squared Error (MSE) and a 43.93% reduction in Mean Absolute Error (MAE), offering a novel perspective for trajectory prediction.