Machine-Learning Algorithm and Decline-Curve Analysis Comparison in Forecasting Gas Production

This study utilizes machine-learning algorithms to reinterpret existing datasets originally plotted using Decline-Curve Analysis (DCA), aiming to enhance predictive accuracy without requiring new field-data acquisition. Historical production records were compiled: monthly oil/gas rates, bottom-hole pressures, and cumulative productions, which were fitted to Arps equations via least-squares optimization, and key decline parameters, such as initial rate, nominal decline rate, and hyperbolic exponent, served as input data. Four machine-learning models were trained and validated: Artificial Neural Networks (ANN), Support Vector Machines (SVM), and Linear Regression (LR), using 80/20 train–test splits and 5-fold cross-validation. Models were evaluated using Mean Squared Error (MSE), Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and coefficient of determination (R2). The ANN emerged as the best-performing method, achieving near-unity predictive accuracy (R2 ≈ 1) on the independent test set, with low error values (MSE = 0.0012 Ncm2/month2, RMSE = 0.035 Ncm/month, MAE = 0.028 Ncm/month) for oil production rates. Similar levels of accuracy were obtained for gas rates and pressures. These results reflect the strong and highly regular relationships present in the dataset analyzed rather than an exact zero-error fit. The multi-layer architecture of the ANN effectively captured the nonlinear interactions between Arps parameters and transient flow regimes, outperforming the empirical and physics-constrained approaches. Linear regression yielded strong results (R2 = 0.98, RMSE = 0.15 Ncm/month) but faltered in high-decline scenarios, failing to model exponential tails accurately. SVM exhibited the highest deviations (RMSE = 0.42 Ncm/month, R2 = 0.89), attributable to kernel sensitivity in sparse, noisy decline data. RF provided intermediate performance (R2 = 0.97). This ANN-driven approach redefines decline analysis by automating parameter tuning and uncertainty quantification, reducing forecasting errors by 85% versus classical Arps methods.