To improve the predictive ability of existing models for predicting the compaction density of asphalt pavement， a test site was set up on the upper layer of the Xiang’an Airport Highway Project in Xiamen. The CCV， DMV and VCV， which represent the change of harmonic ratio， energy， and mechanics in the vibration and compression process， respectively， as well as the temperature， were chosen as indicators. The isolation forest algorithm was used to detect outliers in indicators. The density prediction model was established based on the partial least squares regression. The results show that the isolation forest can effectively recognize outliers of high-dimensional data， covering the shortage that traditional methods can only process one-dimensional data. There are different degrees of positive correlation between temperature， other indicators， and asphalt pavement density. The multiple regression model based on CCV， DMV， and VCV obtains better fitting ability than the unitary regression methods， proving the feasibility of multiple indicators. The partial least squares regression can restrain the adverse impact caused by the approximate collinearity between independent variables， correct the incorrect weight of temperature， and improve the fitting degree compared with the common multiple linear regression methods. The final determination coefficient of the model on the training set is 0.83， and on the test set is 0.81， indicating good predictive ability for asphalt pavement density.