Prediction of Available Phosphorus Formation and Analysis of Key Influencing Factors during Organic Waste Composting Using Stacking Ensemble Learning

• • The stacking ensemble model (XGBoost + SVR base learners, ElasticNet meta-learner) achieved R² = 0.928, RMSE = 0.958, and MAE = 0.691 on the test set, significantly outperforming single models, indicating robust generalization for AP prediction in composting. • • SHAP global importance ranked feedstock type, bulking agent type, turning interval, pH, EC, and C/N as the top six factors; livestock manure and straw-based bulking agents showed significant positive contributions to AP, guiding feedstock selection for phosphorus-rich compost. • • PDP analysis revealed that maintaining pH and C/N at relatively low levels throughout composting favors AP accumulation; specifically, during the cooling/maturation phase, moisture <48% and C/N <14 with composting duration >43 days enhanced AP, providing operational thresholds for process optimization. • • The model's predictive accuracy (R² > 0.92) enables rapid estimation of AP without lengthy experiments, reducing time and cost in composting process design and enabling real-time adjustments to maximize phosphorus availability.