Explainable feature-driven optimization using interpretable machine learning: Insights from concrete temperature control

Temperature control of the concrete surrounding the spiral case (SCSC) is essential for preventing thermal cracking and ensuring long-term structural reliability in large hydropower plants. This study develops an interpretable knowledge-guided machine learning framework for predicting key thermal indicators and supporting cooling-parameter design during construction. The framework employs Six machine learning methods to model the nonlinear thermal behavior using ten variables that characterize structural configuration, construction conditions, and water-pipe cooling settings, with peak temperature and daily temperature-drop rate selected as the target outputs. To enhance model transparency and provide actionable guidance, Shapley Additive Explanations (SHAP) are used to quantify the contribution of each input feature at both global and local scales. Building on this interpretability, an optimization model is further constructed in which SHAP-derived feature importance serves as adaptive weights, enabling the inversion and refinement of cooling-pipe parameters in a physically meaningful manner. Application results show that the proposed approach not only improves the accuracy and interpretability of thermal prediction but also facilitates explainable and efficient temperature-control decision-making for complex concrete structure. The proposed approach integrates model interpretation with optimization, providing a practical and explainable solution for temperature control design in complex concrete structures.