Aiming at the problems that the random forest (RF) used in predicting compaction quality of core gravel soil, such as decision tree number selection and ignoring the influence of P0.075 mass fraction on compaction quality, a random forest (FOA-RF) algorithm based on the fruit fly optimization algorithm (FOA)is proposed, and a core wall gravel soil compaction quality prediction model based on the FOA-RF algorithm considering the content of P0.075 is constructed.On one hand, this model can analyze the correlation between material source parameters and dry density, and P0.075 content can be added as the input parameter.On the other hand, the FOA algorithm is used to optimize the random forest, which solves the problem that the RF algorithm is difficult to obtain the optimal solution of decision tree number and does not consider the influence of decision tree number and random feature number at the same time. Finally, taking a gravel-core wall rockfill dam project in construction in Southwest China as an example, the prediction model based on the traditional RF algorithm, BP neural network, multiple linear regression and the FOA-RF model was used to predict the compaction quality respectively. The result shows that the FOA-RF algorithm has superiority in prediction accuracy. Based on this model, a compaction quality prediction module can be developed and embedded in a real-time monitoring system for rolling quality, which can realize real-time prediction of compaction quality.