Foundation-Specific Hybrid Models for Expansive Soil Deformation Prediction and Early Warning

Foundation deformation prediction on expansive soils involves complex soil-structure interactions and environmental variability. This study develops foundation-specific hybrid modeling approaches for temporal deformation prediction using 974 days of monitoring data from four foundations on medium-expansive soil. Four hybrid architectures were evaluated—Residual-Clustering Hybrid, Elastic Net Fusion, Residual Correction, and Enhanced Robust Huber—optimized through Ridge regression-based feature selection and validated against seven baseline methods. Systematic feature engineering with optimal selection identified foundation-specific complexity requirements. Statistical validation employed bootstrap resampling, temporal cross-validation, and Bonferroni correction for multiple comparisons. Results demonstrated foundation-specific effectiveness with distinct hybrid model performance: Residual-Clustering Hybrid achieved optimal performance for Foundation F1 (R2 = 0.945), Elastic Net Fusion performed best for Foundation F2 (R2 = 0.947), Residual Correction excelled for Foundation F3 (R2 = 0.963), and Enhanced Robust Huber showed strongest results for Foundation F4 (R2 = 0.881). Statistical significance was achieved in 35.7% of comparisons with effect sizes of Cohen’s d = 0.259–1.805. Time series forecasting achieved R2 = 0.881–0.963 with uncertainty intervals of ±0.654–0.977 mm. Feature analysis revealed temporal variables as primary predictors, while domain-specific features provided complementary contributions. The early warning system achieved F1-scores of 0.900–0.982 using statistically derived thresholds. Foundation deformation processes exhibit strong autoregressive characteristics, providing enhanced prediction accuracy and quantified uncertainty bounds for operational infrastructure monitoring.