Forecasting Thermospheric Density with Transformers for Multi-Satellite Orbit Management

TL;DR

This work addresses the challenge of forecasting thermospheric density for Low Earth Orbit operations, where accurate density predictions directly impact satellite drag and collision risk management. It introduces a transformer-based surrogate that ingests a compact set of exogenous features (e.g., GOES X-ray, OMNI2 proxies) and predicts density up to a three-day horizon, either directly or as a residual correction to the NRLMSIS-2.1 baseline. Across real-data experiments, the transformer markedly improves over a persistence baseline on multiple metrics, with the end-to-end approach generally delivering the lowest absolute errors and the residual approach yielding smoother, more stable predictions. The model’s practical significance lies in providing a fast, droppable-density forecast tool suitable for mission planning and multi-satellite orbit management, while acknowledging current limitations in handling spontaneous solar events and the need for larger datasets and uncertainty quantification in future work.

Abstract

Accurate thermospheric density prediction is crucial for reliable satellite operations in Low Earth Orbits, especially at high solar and geomagnetic activity. Physics-based models such as TIE-GCM offer high fidelity but are computationally expensive, while empirical models like NRLMSIS are efficient yet lack predictive power. This work presents a transformer-based model that forecasts densities up to three days ahead and is intended as a drop-in replacement for an empirical baseline. Unlike recent approaches, it avoids spatial reduction and complex input pipelines, operating directly on a compact input set. Validated on real-world data, the model improves key prediction metrics and shows potential to support mission planning.

Figures (3)

• Figure 1: Prediction of the thermospheric density using our proposed transformer with end-to-end and residual learning approaches compared to the persistence baseline and the measured density by GRACE-2 (July, 2013).
• Figure 2: Prediction of the thermospheric density using our proposed transformer with end-to-end and residual learning approaches compared to the persistence baseline and the measured density by SWARM-A (March 2016). Our transformer trained with both flavors is able to capture early events in rising densities.
• Figure 3: Prediction of the thermospheric density using our proposed transformer with end-to-end and residual learning approaches compared to the persistence baseline and the measured density by SWARM-A (August 2019). All models, including the persistent baseline, are unable to capture the rise of the measured density.