| dc.description.abstract | Estimating reserves—economically recoverable volumes of hydrocarbons in a company’s portfolio—requires forecasting hydrocarbon production, which is prone to significant uncertainty and bias. Accurately quantifying this uncertainty is paramount to estimators understanding risk and projects meeting expectations.
Typically, production forecasts are made deterministically using Decline Curve Analysis (DCA). However, production forecasts can also be created probabilistically using Probabilistic Decline Curve Analysis (PDCA). In recent years, some reserves evaluators have turned to multivariate Machine Learning (ML) models to perform deterministic production forecasts, due to ML models’ ability to handle large datasets and include properties other than production in the forecast. However, these models are deterministic and, to the best of my knowledge, there has been no standalone probabilistic adaptation published in the petroleum literature as of yet.
The aims of this research were to determine if a ML method was probabilistically reliable in forecasting production and to determine if the accuracy, probabilistic reliability, predicted uncertainty, and computational cost of this method was superior to an existing PDCA method.
A Gradient Boosting Regressor (GBR) was adapted to generate cumulative production predictions by training three separate models for each of the 10%, 50% and 90% quantiles. Predictions were made with this Gradient Boosting Regressor with Quantiles (GBRQ) method for future months based on the first 12 months of cumulative production history for the training wells, the target cumulative production at the forecasted month for the training wells, and the first 12 months of cumulative production history for the test wells.
Prediction accuracy was measured using the root mean square error (RMSE) between the predicted median (P50) and true values as well as between the predicted mean and true values. Probabilistic reliability was assessed using calibration plots in which the frequency with which actual production values were less than predicted production values at each quantile was plotted against the assigned probability. Predicted uncertainty was assessed using an average normalized uncertainty window and cost was compared on the basis of computational time.
The GBRQ method was more accurate at late times, was more probabilistically reliable, predicted less uncertainty, and was less computationally intensive than a published Probabilistic Decline-Curve-Analysis (PDCA) method for a dataset consisting of 438 conventional wells in the Midland Basin.
The GBRQ methodology can be useful to three groups: (1) reserves estimators, who can make point estimates and full forecasts of probabilistic production comparatively fast and with probabilistic reliability for large datasets; (2) reserves auditors, who can quickly use this method to compare with an auditee’s probabilistic production forecast; and (3) investors and banks, who can evaluate asset acquisitions and divestitures with well-calibrated probabilistic production forecasts. | |
