Asset Development and Sweet Spot Identification in Unconventional Reservoirs Using Machine Learning Approaches
Despite the successful development of unconventional oil and gas resources in the U.S, the industry quickly realized that not all unconventional assets are viable targets for economic hydrocarbon production, and operators sought technologies that could identify appropriate targets for development. Unconventional formations that offer the best potential require a unique combination of reservoir and geomechanical rock properties; such formations are relatively rare. Extremely small pore size and ultralow matrix permeability, make these unconventional resource plays fundamentally different from most conventional reservoirs. Furthermore, because hydrocarbon migration paths tend to be short, and productive zones of shale reservoirs may be confined to a certain area within a basin or restricted to a stratigraphic interval. This implies seeking viable economic unconventional plays is not enough. As such it is critical to associate with unconventional reserves, frackability and producibility. Therefore, designing optimum exploitation programs for unconventional plays requires a multi-faceted reservoir and production characterization scheme which can simultaneously address several issues. Currently, unconventional play extraction processes are quite inefficient. The industry is in a state of trial-and-error, with only immature research results available regarding the matching of choicest unconventional plays, with peak exploration, drilling, completion, and production programs. This implies optimizing asset development and identifying ‘Sweet Spots’ can help boost efficiency. Throughout this study, we have attempted to use machine learning approaches to identify ‘Sweet Spots’ (the most prolific locations/wells) propose new ways to improve/optimize unconventional asset development. To begin with, we put forward a workflow that allows for the identification of areas within a given unconventional asset that have suitable reservoir and completion characteristics. The workflow leverages the power of recurrent neural networks to generate synthetic well log data (where necessary), which is then combined with existing real log data, and used as a basis (inputs) for sequential gaussian simulation. By so doing, we were able to utilize readily available well log data to pinpoint promising regions of the upper Bakken. In addition, we proposed a neural net capable of predicting the cumulative 6 months of production for a given well using various completion and reservoir parameters. This approach improves our ability to gauge a well’s potential productivity with varying completion scenarios. As such we can maximize the productivity from each. Finally, we introduced a novel Generative Adversarial framework (AECGAN) capable of forecasting long-term oil production rates. This AECGAN utilizes spectrograms to accommodate historical data of varying durations. With this forecast in hand, engineers can evaluate the trend of various wells and intervene, when necessary, thus boosting the productivity of each well.