Accurate prediction of the average thermal extraction load (ATEL) in hydrothermal heating systems optimizes energy recovery, though numerical models are constrained by modeling accuracy and computational time costs. Data-driven modeling can efficiently screen alternative production plans and enhance daily operational decisions. However, ensemble learning applications for forecasting ATEL lack comprehensive comparisons and appropriate algorithm selection. This study evaluates six ensemble learning algorithms: random forest (RF), adaptive boosting (AdaBoost), gradient boosting decision trees (GBDT), extreme gradient boosting (XGBoost), light gradient boosting machine (LightBGM) and categorical boosting (CatBoost) on 1159 data points from a homogeneous thermal-hydraulic coupling numerical model. The superior performance of the CatBoost is characterized by the highest accuracy (with the lowest root mean square error (RMSE) of 150.6 kW–170.7 kW, minimal mean absolute percentage error (MAPE) of 0.66%–0.73%, and exceptional R2 values of 0.999), commendable stability, reasonable computational cost and low uncertainty. LightGBM, XGBoost, and GBDT are identified as suitable substitutes, whereas RF is a possible substitute. Moreover, the production rate is identified as the most significant factor contributing to the ATEL, followed by the well spacing and injection temperature. While perforation depth has the least influence on heat recovery.