Abstract: Modern engineering increasingly demands high-performance building materials. Steel fiber-reinforced concrete (SFRC), a composite material with excellent mechanical properties and durability, has therefore gained wide application. The compressive strength of SFRC is a key indicator of its performance; however, conventional laboratory testing through standardized curing and loading procedures requires significant manpower, costly materials, and extended curing times that often span several weeks. To address these limitations, this study proposes an innovative predictive framework based on stacking ensemble learning to efficiently and accurately estimate SFRC compressive strength. A comprehensive dataset comprising 211 distinct SFRC mix proportions was collected and randomly divided into training (169 samples) and testing (42 samples) sets using an 8:2 ratio. Principal component analysis (PCA) reduced the dimensionality of the original 10 input variables to 5 principal components, thereby simplifying computation and mitigating risks of overfitting. Four machine learning algorithms—Support Vector Machine (SVM), Decision Tree (DT), K-Nearest Neighbors (KNN), and Random Forest (RF)—were employed as base learners, while a Backpropagation Neural Network (BP) served as the meta-learner in the stacking framework. To further improve predictive accuracy, six optimization algorithms—Whale Optimization Algorithm (WOA), Sparrow Search Algorithm (SSA), Gray Wolf Optimizer (GWO), Arithmetic Optimization Algorithm (AOA), Chaos Game Optimization (CGO), and Particle Swarm Optimization (PSO)—were applied to tune hyperparameters of the individual models. The optimally tuned models were then integrated through a stacking ensemble strategy, producing the OP-Stacking hybrid model. The predictive performance of OP-Stacking was evaluated against the five individual models using both the training and testing datasets. Results demonstrated that OP-Stacking significantly outperformed its component models across all evaluation metrics. On the training set, OP-Stacking achieved a mean squared error (MSE) of 29.9903, mean absolute error (MAE) of 3.621, mean absolute percentage error (MAPE) of 8.39%, and coefficient of determination (R²) of 0.9772. On the testing set, the model achieved an MSE of 96.4937, MAE of 7.1328, MAPE of 12.99%, and R² of 0.9332. Comparative analysis confirmed the superiority of OP-Stacking in terms of predictive accuracy, generalization ability, and fitting performance. Beyond short-term prediction, linear regression analysis was conducted to establish a relationship between the 7-day and 28-day compressive strengths of SFRC. The analysis yielded a strong correlation with R² = 0.963 and produced an empirical formula describing the development of compressive strength from early to later curing stages. This empirical relationship enhances the practical utility of the proposed framework by enabling reliable long-term strength predictions from short-term data. Finally, the OP-Stacking hybrid model and the derived empirical formulas were encapsulated within a user-friendly predictive system developed using the Qt Framework. This integrated tool supports both compressive strength prediction and intelligent mix proportion design for SFRC. The system has already demonstrated its value by providing essential technical support for the accelerated construction of innovative support structures in the Central Yunnan Water Diversion Project.