Explainable machine learning models to predict outlet water temperature of pipe-type energy pile

To address the interpretability gaps and data scarcity in predicting summer outlet water temperature of pipe-type energy piles, this study proposes a hybrid framework integrating multi-physics simulation and explainable machine learning. A 3D transient heat transfer model was developed in COMSOL to generate 1000 simulation datasets covering key operational parameters (inlet water temperature, water velocity, material thermal properties). Four supervised learning algorithms (KNN, Regression Tree, Random Forest, BPNN) were implemented, with SHAP (Shapley Additive Explanations) for feature contribution quantification. Results show that the BPNN model achieved the highest accuracy (RMSE = 0.448 degrees C), outperforming RF by 32 %. SHAP analysis the relative contributions of inlet water temperature (51.2 % contribution), water velocity (21 %) and material thermal properties (27.8 %). This work provides data-driven insights for pipe-type energy pile optimization, with future extensions planned for multi-size piles and real-time predictive models.