Robust and Interpretable Boosting-Based Ensemble Machine Learning Model for Predicting Production Temperature in a Geopressured Geothermal System of the Yinggehai Basin， South China Sea

Accurate prediction of production temperature （PT） is essential for optimizing the design and operation of geothermal systems. However， most machine learning （ML） studies emphasize accuracy while neglecting statistical significance， uncertainty quantification and interpretability. This study develops a robust and interpretable boosting-based ensemble ML model for PT prediction in a geopressured geothermal system in the Dongfang area of the Yinggehai Basin. A dataset of 581 samples was generated from coupled reservoir—wellbore simulations. Six boosting-based ensemble ML models—AdaBoost， GBDT， XGBoost， LightGBM， CatBoost and NGBoost—were developed and evaluated using conventional metrics， Wilcoxon signed—rank tests， bootstrap-based uncertainty assessment and Shapley Additive exPlanations （SHAP）. XGBoost achieved the best test performance， with a mean absolute percentage error of 0.370%， a root mean square error of 0.619℃ and an R² of 0.998. No significant difference was detected between XGBoost and CatBoost （p =0.058 and |z|=1.896）. Among the three models evaluated by bootstrap analysis， XGBoost showed the lowest prediction variability. SHAP analysis identified well spacing as the most influential feature and quantified global and instance-specific contributions. Overall， XGBoost provides accurate， robust and interpretable PT predictions， offering an efficient surrogate within the investigated design space.