Abstract VLBI data processing contains real-time and after-action calculation modes. In VLBI real-time mode, tropospheric delay correction can provide forecast value, and a high precision tropospheric empirical model is needed to realize the real-time correction of tropospheric delay. Currently, various applications require high-precision tropospheric empirical models to achieve real-time correction of tropospheric delay. In this paper, the IGS zenith tropospheric delay (ZTD) sequence and the ERA5 grid data integration ZTD sequence from 2010 to 2012 are used to establish empirical models on four VLBI sites in China as Local_ERA and Local_ZTD respectively. Based on the 2013-2014 IGS site ZTD, the two improved models are compared with the global GPT2w and SHAO-Gm models. The results show that the improved Local_ERA and Local_ZTD models have similar precision. The average accuracy of the model has improved at four VLBI sites, compared to GPT2w and SHAO-Gm, especially in the Beijing station with strong seasonal variation of water vapor. The average accuracy of Local_ERA is slightly higher than that of Local_ZTD, which is 7.90% higher than that of GPT2w model, and 21.26% higher than SHAO-Gm model.