Interpretable predictive modelling of outlet temperatures in Central Alberta's hydrothermal system using boosting-based ensemble learning incorporating Shapley Additive exPlanations approach

Accurately predicting outlet water temperature is crucial for optimizing the sustainability of geothermal system exploration. However, the practical application of machine learning (ML) in geothermal systems is often limited by a lack of interpretability and transparency, despite their outstanding predictive performance. This study addresses these limitations by employing boosting-based ensemble models—including boosted decision trees, extreme gradient boosting, light gradient boosted machine and category boosting (CatBoost)—integrated with Shapley Additive exPlanations (SHAP) to predict outlet water temperature in Central Alberta's hydrothermal field. Numerical simulation data were used for model training and testing, with sensitivity analysis identifying input features. CatBoost provided the highest prediction accuracy, achieving a root mean square error of 0.278, a mean absolute percentage error of 0.35%, and a coefficient of determination of 0.999, validated through absolute relative percentage error and residual analysis. SHAP analysis identified well spacing as the most influential factor, followed by production rate, horizontal well length, and injection temperature. Feature interactions and non-linear effects were significant, with combinations of features outperforming individual ones. Local SHAP interpretations emphasized the effects of well spacing on specific predictions. These findings improve ML model interpretability and provide actionable recommendations for optimizing geothermal energy management.