Abstract: In order to analze the stochastic characteristics of BDS (BeiDou satellites) observations, GEO(geostationary earth orbit),IGSO(inclined geosynchronous satellite orbit),MEO(medium earth orbit) satellites observations variance components are estimated. The observation variance components of different receivers and different types of BDS are proved to be significantly different in size and time-varying characteristics. To deal with the heterogeneous variances, a real-time estimating BDS observations stochastic model procedure is developed with the methods of iterative least-squares and MINQUE(minimum norm quadratic unbiased estimation).To demonstrate its performance, precise relative positioning experiments are carried out using 243m, 645m and 10 137 m baselines, respectively. In contrast to the empirical stochastic model, the real-time estimating stochastic model can improve the performance of BDS precise relative positioning, and improve more significantly for the longest baseline.10 137 m baseline test results indicate that the real-time estimating stochastic model can improve mean positioning precision 41.3%、54.5%、51.6% and RMS values of multi-epoch position offsets 38.4%、17.7%、39.7% in N(north), E(east), U(up) direction respectively.