The invention relates to a method for modeling random characteristics of multi-frequency GNSS (global navigation satellite system) observed values. The method comprises the following steps of acquiring multi-frequency GNSS observed data, and preprocessing the data; constructing a single difference observation equation to form an intersite single difference observed value; performing parameter reforming on the single difference observation equation according to a base line and fixed double-difference ambiguity; taking the average value of single difference observed values of single-epoch multiple satellites as the least square solution of the reformed parameter, and subtracting the least square solution from the single different observed value of each satellite to obtain single difference observation noise; calculating the accuracy of non-difference observed values of single-epoch multi-frequency GNSS different-type observed values, cross-correlation coefficients of the different-type observed values and temporal correlation coefficients of the same-type observed values by utilizing the extracted observation noise; obtaining a relation between the accuracy of the observed value of each satellite and an elevating angle; modeling, outputting model parameters and establishing a variance-covariance matrix. Compared with the prior art, the method has the advantages of simple calculating process, reliability and the like.