Abstract: Objectives: the predicted Global Navigation Satellite System (GNSS) ultra-rapid clock offset products are the one of time references of high-precision real-time and near-real-time location services. However, the quality of traditional ultra-rapid clock offset products are directly affected by the noise and reliability of the estimated ultra-rapid clock offset series. In this research, an improved GNSS ultra-rapid clock offset predicted model by inserting the frequency stability of the satellite onboard atomic clocks is proposed. Methods: first, the velocity and acceleration terms of clock offset are introducing into its estimation equations to acquire the stabler origin series than the traditionally estimated clock offsets. Second, the velocity and acceleration parameters are predicted based on the combination of long short-term memory (LSTM) and (Multilayer Perceptron) MLP models to fully consider the spatiotemporal correlation of the clock offset state series. Third, the sliding window is used to model the continuously predicted clock offset state series to integrate the epoch increment of predicted clock offsets. Results: according to the experiments, it is indicated that the stability of clock offset velocity and acceleration is significantly improved compared with the traditional ultra-rapid clock offset series. Meanwhile, the accuracy of the predicted ultra-rapid clock offset based on the proposed model can be improved at least 14.0%, 7.9% and 22.3% within 6h, 12h and 24h, respectively. In addition, the recovered predicted ultra-rapid clock offset based on the integration can optimize the positioning accuracy and convergence time of static PPP solutions in E, N and U directions with 74.0%、 34.3%、 51.4% and 46.5%、 42.7%、 63.3%. Conclusions: therefore, the quality of GNSS satellite ultra-rapid clock offset products can be improved based on the proposed model, which will effectively supplement to the rapid precision location service products.