Neural network-based clock bias prediction for low earth orbit satellites with a reconstruction fine-tuning mechanism

Abstract Accurate prediction of Satellite Clock Bias (SCB) for Low Earth Orbit (LEO) satellites is critical for real-time LEO-enhanced Global Navigation Satellite System (LeGNSS) applications. However, the inherent instability of LEO onboard oscillators results in complex SCB variations that are difficult for conventional prediction models to capture. While deep learning methods offer a promising alternative, their effectiveness can be limited by two key issues: the pronounced high-order trends in LEO SCB, which are not completely removed by the conventional single differencing; and the potential biased error accumulation that occurs when reconstructing predictions from the differenced domain. To address these challenges, we propose a novel neural network-based framework featuring two key innovations: (1) advanced differencing strategies—single-differencing with segment-wise standardization and direct double-differencing—to suppress high-order trends without performing external fitting, thereby preserving prediction independence; and (2) a reconstruction fine-tuning mechanism to explicitly mitigate directional error accumulation. Validated with Gravity Recovery and Climate Experiment Follow-On (GRACE-FO) data, the fine-tuning mechanism reduces overall prediction errors by 6%–16% with reconstruction biases mitigated. Compared to the conventional spectrum analysis model, the proposed framework improves overall prediction accuracy by approximately 90% and 94% for the 60 min and the 10 min prediction horizons, respectively. Specifically, for the prediction length of 60 min, the prediction errors for the GRACE-FO C and D satellites decrease to 0.84 ns and 1.42 ns, respectively; for the prediction length of 10 min, they decrease to 0.10 ns and 0.15 ns, respectively. These results indicate the framework’s potential for high-precision real-time applications from the inference-efficiency perspective.