Comparative evaluation of gas well production forecasting using decline curve analysis and data-driven predictive models

Accurate forecasting of gas well performance is essential for optimizing reservoir management and enhancing field development strategies. It enables informed decision-making for production planning, well intervention, and long-term economic evaluation of hydrocarbon reservoirs. So, this study presents a comparative analysis of two forecasting approaches, Decline Curve Analysis (DCA) and Machine Learning (ML), to evaluate their effectiveness in predicting daily gas production. DCA was implemented using both MBAL reservoir simulation and Python-based curve fitting with SciPy, while the ML approach employed Gradient Boosting Regressor (GBR), Light Gradient Boosting Machine (LightGBM), and AdaBoost algorithms. A Monte Carlo simulation was used to quantify uncertainty in cumulative production forecasts, and a Taylor diagram assessed model performance in terms of correlation, standard deviation, and root mean square error. Among all tested models, the LightGBM algorithm achieved the best performance with an R² of 0.9845, RMSE of 0.3672, MAE of 0.1232, and PBIAS of 0.05% on the testing dataset, demonstrating strong predictive capability and minimal systematic bias. MBAL also performed reliably but required detailed reservoir input data and homogeneity assumptions, whereas SciPy-based curve fitting showed greater early-stage deviations. The Monte Carlo results confirmed the robustness of all models, with the LightGBM P50 forecast closely matching observed production trends and uncertainty bounds. The integration of analytical and data-driven methods offers a unified framework for more accurate and uncertainty-aware gas production forecasting. The findings underscore the practical advantage of ensemble ML models, particularly LightGBM, as efficient, assumption-light, and scalable tools for petroleum reservoir performance prediction.