Abstract: The nonlinear characteristics of satellite clock offset sequence are complex, in order to ef fectively separate the correction error of periodic term and take into account the factors that cannot be functionalized, and improve the accuracy of clock bias prediction, the clock error periodic term model is extended to the semi-parametric clock error model. By using kernel estimation method, the window width parameter and parameter solution of kernel function are comprehensively considered, a semi-parametric varying coefficient model is established. Firstly, the kernel estimation method is i ntroduced, the parameter components are modified by Taylor expansion, by synthesizing kernel funct ion and window width parameters, the parameters and periodic term correction estimates of the semi -parametric variable coefficient model are obtained by three-step estimation method; Then, score test statistics are constructed to identify outliers, and a method of identifying outliers is proposed for t he semi-parametric variable coefficient clock difference prediction model with periodic correction; Fi nally, in order to avoid over-fitting or under-fitting of the observed values, support vector machine i s used to further fit the pre-processed clock residual data to improve the fitting and prediction accu racy of the model. The clock difference data of BDS-3 satellite is compared with the conventional methods to verify the reliability of the new method. Experimental results show that the proposed m ethod can accurately and efficiently locate the constant value of BDS-3 clock difference, identify an d separate periodic item corrections, and greatly improve the quality and efficiency of BDS clock d ifference data preprocessing. The prediction accuracy of the combined forecast model in this paper i s better than that of the traditional quadratic polynomial model, periodic term model and semi-para metric model. For 1h, 6h and 12h forecast, the average accuracy of the forecast of BDS-3 satellite clock difference data is better than 0.1635ns.