In order to address the long testing cycle and the difficulty of empirical models in capturing complex nonlinear interactions among multiple factors in predicting the 28-day compressive strength of concrete, a prediction framework integrating multi-model optimization and multiscale interpretability was developed in this study. Based on the UCI dataset, six key input variables were selected using LASSO-CV regularization, random forest MDI feature importance, and variable ablation experiments. The selected variables were cement, water, blast furnace slag, fly ash, superplasticizer, and coarse aggregate. Four models, namely K-Nearest Neighbors (KNN), Support Vector Regression (SVR), Random Forest (RF), and Gradient Boosting Regression (GBR), were constructed and optimized using Particle Swarm Optimization (PSO), African Vulture Optimization Algorithm (AVOA), Bayesian Optimization, and Genetic Algorithm (GA), respectively. A stacking ensemble model was then developed using Ridge Regression as the meta-learner and an out-of-fold prediction strategy. SHAP, Partial Dependence Plots (PDP), and LIME were further integrated for multiscale interpretability analysis. The results show that the stacking model achieves the best performance on the test set, with an R2 of 0.943 and an RMSE of 3.613. Furthermore, 30 repeated stratified 7:3 random split experiments show that the stacking model achieves the highest average R2 (0.8306 ± 0.0265) and the lowest average RMSE (5.6619 ± 0.4934) among all compared models, indicating good generalization performance and stability. SHAP analysis identifies cement and water as the dominant variables affecting the 28-day compressive strength of concrete. PDP illustrates the main effects and interactions among the key variables, while LIME provides additional local interpretations. An intelligent prediction and interpretation system is constructed to support engineering applications. The proposed method provides an effective tool for concrete mix proportion optimization and rapid strength evaluation, with high prediction accuracy and favorable interpretability.