INTERPRETABLE MACHINE LEARNING FOR GEOTHERMAL POWER GENERATION PREDICTION: AN XGBOOST–SHAP APPROACH

Abstract

Geothermal power plants require accurate forecasting models to maintain operational stability and improve energy efficiency. However, electricity generation is influenced by complex thermodynamic and chemical parameters that are difficult to model using conventional approaches. Therefore, this study proposes an explainable machine learning frame work based on Extreme Gradient Boosting (XGB) and SHapley Additive exPlanations (SHAP) for electricity generation prediction using geothermal operational data. The dataset consisted of 14,252 hourly operational records collected from March 2022 to December 2024. Model performance was evaluated using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and coefficient of determination (R2), while SHAP was employed to analyze feature contributions and improve model interpretability. Experimental results showed that the XGB model achieved an MAE of 0.26009 MW, an RMSE of 0.47537 MW, and an R2 of 0.99499. SHAP analysis revealed that steam flow rate and pressure-related variables were the most influential factors affecting electricity generation prediction. These findings demonstrate that the proposed XGB-SHAP framework can provide both accurate prediction performance and interpretable insights for geothermal power plantoperation and performance monitoring.