A field survey was first conducted to collect merging information of on-ramp vehicles at urban expressway. Then clustering analysis was used to divide merging maneuvers into three clusters: the forepart, the midpiece and the endpiece. Five distribution models were selected to fit the distribution of merging positions in each cluster. Model parameters were determined by maximum likelihood estimation, and the goodness of fit to these models was also given by Kolmogorov-Smirnov test. It was found that the Johnson SB model obtains the best result in the forepart cluster, the Log-logistic model is the most suitable for the midpiece merging maneuvers, and the Gen. extreme. value model provides the best fit for the endpiece cluster. At last, a regression model was proposed to describe the relationship among the distribution of merging positions, the density of the shoulder lane, the merging velocity of an on-ramp vehicle. The result of sensitivity analysis shows that the density and the velocity have significant effect on the distribution of merging position in the same cluster.