Aiming at the requirements of the real-time process evaluation for the on-site compaction degree of roller-compacted concrete(RCC), prediction parameters such as the moisture content, the shear wave velocity on the surface of compacted fresh concrete, aggregate gradation, the ratio of cementitious material and sand are selected to construct a prediction model based on GA-BP neural network. The effectiveness of the proposed real-time evaluation method was validated by an on-site application case. The results show that compared with BP neural network models, GA-BP model has a higher prediction accuracy with a smaller range of deviation fluctuation and it can predict the real-time compaction degree of RCC accurately and effectively with higher stability. GA-BP model is more sensitive to the lower limit value of the compaction degree. For samples with the compaction degree ranging from 93% to 96%, the average error of the GA-BP model is only 0. 08% and the maximum one is only 0. 17%, revealing very high prediction accuracy.