Summary
Following the rapid growth of unconventional resources, many models and methods have been proposed for forecasting the performances of unconventional wells. Several studies have attempted to use machine learning (ML) for improving the forecasting. However, owing to limitations of ML in regard to long-term forecasts (e.g., the occurrence of unphysical results), most of these ML forecasts are not satisfactory. In this work, we propose, demonstrate, and discuss a new ML approach able to rapidly provide probabilistic, long-term forecasts of oil production rates from individual wells in a decline curve analysis (DCA) manner. The novelties of the proposed approach are as follows: (1) it combines an automated ML (AutoML) method for supervised learning and a Bayesian neural ordinary differential equation (BN-ODE) framework for time-series modeling; (2) it uses the DCA model to inform the BN-ODE framework of “physics” and regulate the BN-ODE forecasts; and (3) several completion parameters (such as locations, lengths, and slickwater volume) of individual wells are analyzed and included as the inputs of model building, in addition to measured oil production rate data. Specifically, AutoML method is first used to model the relationship between the well location, completion parameters, and the DCAs parameters, and the BN-ODE framework is then used to model the relationship between the DCAs parameters and the time-series oil production rates. A publicly accessible data set, consisting of completion parameters and oil production rates, of 396 horizontal wells in the Bakken Shale Formation is used to train and test the model of the proposed approach. The results lead to the conclusion that the proposed approach is practical for providing probabilistic, long-term forecasts of oil production from individual wells, given data of existing wells in the reservoir.
