Analysis centers (ACs) for global navigation satellite systems (GNSSs) cannot accurately obtain real-time Earth rotation parameters (ERPs). Thus, the prediction of ultra-rapid orbits in the international terrestrial reference system (ITRS) has to utilize the predicted ERPs issued by the International Earth Rotation and Reference Systems Service (IERS) or the International GNSS Service (IGS). In this study, the accuracy of ERPs predicted by IERS and IGS is analyzed. The error of the ERPs predicted for one day can reach 0.15. mas and 0.053. ms in polar motion and UT1-UTC direction, respectively. Then, the impact of ERP errors on ultra-rapid orbit prediction by GNSS is studied. The methods for orbit integration and frame transformation in orbit prediction with introduced ERP errors dominate the accuracy of the predicted orbit. Experimental results show that the transformation from the geocentric celestial references system (GCRS) to ITRS exerts the strongest effect on the accuracy of the predicted ultra-rapid orbit. To obtain the most accurate predicted ultra-rapid orbit, a corresponding real-time orbit correction method is developed. First, orbits without ERP-related errors are predicted on the basis of ITRS observed part of ultra-rapid orbit for use as reference. Then, the corresponding predicted orbit is transformed from GCRS to ITRS to adjust for the predicted ERPs. Finally, the corrected ERPs with error slopes are re-introduced to correct the predicted orbit in ITRS. To validate the proposed method, three experimental schemes are designed: function extrapolation, simulation experiments, and experiments with predicted ultra-rapid orbits and international GNSS Monitoring and Assessment System (iGMAS) products. Experimental results show that using the proposed correction method with IERS products considerably improved the accuracy of ultra-rapid orbit prediction (except the geosynchronous BeiDou orbits). The accuracy of orbit prediction is enhanced by at least 50% (error related to ERP) when a highly accurate observed orbit is used with the correction method. For iGMAS-predicted orbits, the accuracy improvement ranges from 8.5% for the inclined BeiDou orbits to 17.99% for the GPS orbits. This demonstrates that the correction method proposed by this study can optimize the ultra-rapid orbit prediction.