Robust Short‐Term Forecasting of Thermospheric Density and Orbit Prediction Using an Interpretable iTransformer Model

Accurate short‐term forecasting of thermospheric mass density is crucial for satellite operations. However, achieving high‐precision predictions remains a significant challenge. In this study, we develop an iTransformer framework that forecasts thermospheric density over the next 24 hr using observations from the preceding 24 hr. Using density data from multiple low Earth orbit satellites (CHAMP, GRACE‐A, GRACE‐C, GOCE and Tianmu‐1/TM02), we investigate two strategies: directly predicting density values and learning the residuals relative to the MSISE‐00 empirical baseline. Performance assessments against independent observations reveal that the residual‐based strategy (iTrans‐Res) generally provides higher accuracy than the direct forecasting approach (iTrans‐Dir), the original and bias‐corrected MSISE‐00 (MSISE‐Res) and a residual‐based Bi‐LSTM (Bi‐LSTM‐Res) baseline. For CHAMP in 2008, iTrans‐Res reaches R 2 = 0.9052 and reduces MAPE from 77.66% to 11.56%; for GRACE‐C in 2024, it reaches R 2 = 0.9432 and reduces MAPE from 38.77% to 9.58%. Storm‐event analyses show that iTrans‐Res more closely reproduces the enhanced density peaks and substantially reduces the recovery‐phase bias inherited from MSISE‐00. Variate‐attention analysis reveals model‐internal associations among the spatiotemporal and space‐weather inputs, with geomagnetic ap information broadly integrated into the variable representations. Orbit propagation tests demonstrate that iTrans‐Res reduces the 24‐hr maximum position error by 24.6% for CHAMP and 16.3% for TM02. Robustness tests with noise‐injected inputs confirm that iTrans‐Res maintains reliability even under moderate space weather forecasting uncertainties.