Abstract We propose a BDS-3 ultra-rapid clock offset prediction algorithm based on EM algorithm optimized relevance vector machine. First, we use the combined MAD method to preprocess the clock data and perform one time difference. Then, we use the one time difference data to train the RVM model, we use the EM algorithm to iteratively obtain the hyperparameters of the model, and finally we use the optimized RVM model to predict. We then restore the one-time difference prediction value of the clock offset to obtain the prediction value of the clock offset. The prediction test is carried out with the measured BDS-3 ultra-rapid clock offset data provided by iGMAS, and the prediction results of this method are compared with the QP model, the SA model and the ultra-rapid clock offset prediction product (ISU-P) of iGMAS. The results show that the mean accuracy of the BDS-3 satellite clock offset data prediction by the RVM model is better than 0.61 ns regardless of whether the forecast is 6 h, 12 h or 24 h. Compared with ISU-P, QP model, SA model, the prediction accuracy of the 24 h BDS-3 satellite clock offset has been improved by 64.1%, 50.0%, 49.2%,respectively.
