Combining semiparametric and machine learning approaches for short-term prediction of satellite clock bias

Accurate modeling of satellite clock bias (SCB) is critical for enhancing high-precision positioning capabilities. Existing approaches, such as semiparametric adjustment models and neural networks, address the nonlinearity and non-stationarity of SCB time series, as well as potential distortions from trend and noise component overlap. However, these methods encounter practical limitations, particularly in the selection of kernel functions for semiparametric models and the initialization of parameters for neural networks. To overcome these challenges, this paper introduces a novel integrated model called the Semi-LFA-Informer (SLFAI) model. Moreover, this model combines semiparametric techniques with optimized self-attention neural networks and is applied to predict SCB for BDS-3. Its performance is compared with other models, including quadratic polynomial (QP), spectral analysis (SA), and long short-term memory (LSTM) networks. The comparison is focused on prediction stability and accuracy. The experimental results show that the proposed method can not only effectively solve the problem of the generalization ability, but also significantly enhance the computational efficiency and accuracy. The SLFAI model achieves average prediction accuracies exceeding 0.15 ns, 0.25 ns, and 0.35 ns for 3-hour, 6-hour, and 12-hour forecasts, respectively, Meanwhile, compared with the other three models, The SLFAI model shows an average prediction accuracy improvement of approximately 53.6%, 59.4%, and 43.5% for the 3-hour, 6-hour, and 12-hour forecasts, respectively, representing a new approach to acquiring high-quality SCB.