Machine Learning-Assisted History Matching in Improving Production Forecast Accuracy for Some Wells in the Rio Del Rey Basin, Cameroon

Abstract Accurate production forecasting is critical for reservoir management and economic evaluation, yet conventional history-matching methods are subjective and time intensive, which can increase uncertainty in the Rio Del Rey (RDR) Basin. This study develops an ML-enhanced history-matching framework to improve production forecast accuracy for wells in the basin. Using Python in a Jupyter Notebook, production data from Well-101S5D and Well-102 were preprocessed, including the imputation of 39 missing values for Well-101S5D. Data were split into 80% training and 20% testing sets. The following ML models were trained and evaluated—Linear Regression (LR), Random Forest Regressor (RFR), XGBoost, Prophet, and LSTM—and their forecasts were compared with decline curve analysis (DCA) models. The selected models were RFR for Well-101S5D and LR for Well-102. ML predictions produced EUR estimates of 0.3269 MMbbl (Well-101S5D) and 0.1274 MMbbl (Well-102). For Well-101S5D, RFR outperformed DCA models (RFR: MSE = 3766.65, MAE = 36.83, R2 = 0.6901; DCA: MSE = 750740.32, MAE = 630.74, R2 = 0.3948). For Well-102, the hyperbolic DCA achieved a strong R2 (0.8055), while ML provided a more consistent and automated workflow. Overall, integrating ML into history matching improves forecast accuracy while increasing objectivity, repeatability, and efficiency. The proposed workflow provides a practical and scalable approach for other complex basins where data-driven history matching is not widely documented.