Satellite clock error prediction plays a key role in GNSS technology for Positioning, Navigation and Timing (PNT) service. However, due to the different types and quality of the space-borne atomic clocks, it is rather difficult to predict high-accuracy satellite clock errors by using one unique prediction model. In general, the performance of existing prediction models varies with different types of satellite clocks, lengths of fitting arcs and predicting arcs. In this article, three common-used prediction models e.g. polynomial model, grey model, and Auto-Regressive Integrated Moving Average (ARIMA) model are evaluated with GNSS precise satellite clock products provided by GFZ. The prediction precision of these models is calculated with respect to precise clock products, specified by different lengths of fitting arcs and prediction arcs respectively. Then, the clock errors predicted by these models are applied to kinematic Precise Point Positioning (PPP) separately and the corresponding positioning performance is discussed. The numerical results show that polynomial model performs the best compared to other two models. For 5-min prediction, the RMS values of predicted clock for GPS, Galileo, GLONASS, and BDS are 0.14 ns, 0.016 ns, 0.21 ns and 0.089 ns respectively. For kinematic PPP with 5-min predicted clock products, polynomial model has the minimum RMS values among three models, with 0.031 m, 0.026 m, and 0.068 m for east, north, and up components, respectively. Moreover, sub-decimeter can be reached for horizontal component with 1-hour predicted clock products. ARIMA model is comparable to polynomial model when the prediction arc is 5–15 min but becomes relatively worse with the increase of the length of prediction arc. Grey model is the worst among the three models, it can meet the requirement of sub-decimeter precision in kinematic PPP only if the prediction arc is shorter than 15 min.

卫星时钟误差预测在用于定位、导航和授时 (PNT) 服务的 GNSS 技术中起着关键作用。然而，由于星载原子钟的类型和质量不同，使用一种独特的预测模型来预测高精度卫星钟误差是相当困难的。一般而言，现有预测模型的性能因卫星时钟类型、拟合弧长和预测弧长的不同而不同。本文利用GFZ提供的GNSS精密卫星时钟产品，对多项式模型、灰色模型、自回归综合移动平均(ARIMA)模型等三种常用的预测模型进行了评估。这些模型的预测精度是针对精密时钟产品计算的，分别由不同长度的拟合弧和预测弧指定。然后，将这些模型预测的时钟误差分别应用于运动学精确点定位（PPP），并讨论了相应的定位性能。数值结果表明多项式模型与其他两种模型相比表现最好。对于5分钟预测，GPS、伽利略、GLONASS和BDS的预测时钟RMS值分别为0.14 ns、0.016 ns、0.21 ns和0.089 ns。对于具有 5 分钟预测时钟产品的运动学 PPP，多项式模型在三个模型中具有最小的 RMS 值，东、北和上分量分别为 0.031 m、0.026 m 和 0.068 m。此外，对于具有 1 小时预测时钟产品的水平分量，可以达到亚分米。ARIMA 模型在预测弧为 5-15 min 时与多项式模型相当，但随着预测弧长度的增加而变得相对较差。格雷模型是三种模型中最差的，只有预测弧短于15分钟才能满足运动学PPP亚分米精度的要求。