Pavement performance prediction is necessary for road maintenance and repair (M&R) management and plans. The accuracy of performance prediction affects the allocation of maintenance funds. The international roughness index (IRI) is essential for evaluating pavement performance. In this study, using the road pavement data of LTPP (Long-Term Pavement Performance), we screened the feature parameters used for IRI prediction using the mean decrease impurity (MDI) based on random forest (RF). The effectiveness of this feature selection method was proven suitable. The prediction accuracies of four promising prediction models were compared, including Gradient Boosting Decision Tree (GBDT), eXtreme Gradient Boosting (XGBoost), support vector machine (SVM), and multiple linear regression (MLR). The two integrated learning algorithms, GBDT and XGBoost, performed well in prediction. GBDT performs best with the lowest root mean square error (RMSE) of 0.096 and the lowest mean absolute error (MAE) of 6.2% and the coefficient of determination (R²) reaching 0.974. However, the prediction accuracy varies in numerical intervals, with some deviations. The stacking fusion model with a powerful generalization capability is proposed to build a new prediction model using GBDT and XGBoost as the base learners and bagging as the meta-learners. The R², RMSE, and MAE of the stacking fusion model are 0.996, 0.040, and 1.3%, which further improves the prediction accuracy and verifies the superiority of this fusion model in pavement performance prediction. Besides, the prediction accuracy is generally consistent across different numerical intervals.