With hundreds of rigs running and thousands of wells producing in unconventional plays, more and more data becomes available every day and it is ever more tempting to apply machine learning techniques for unconventional development, be it to identify geology sweet spot, understand performance drivers and optimize development strategies such as well spacing, completion and production designs etc. However, most of the previous applications of machine learning are limited to either certain types of data or small areas of interest. Consequently, the results often lack the predictability or generalizability necessary to impact important development decisions. We developed a flexible, scalable and integrated machine learning framework to leverage all sources of data for the goal of optimizing unconventional resources development.
The framework is built on a big data warehouse and on-demand capability to efficiently visualize and analyze large volumes of heterogeneous data. The most important pillar of the framework is the ability to transform all different types of data with fit-for-purpose methodologies to be closely related to the evaluation and prediction of well performance. This is enabled mechanistically by an interface to scripting languages such as R or Python for interactive data processing, validation and visualization. We also developed several innovative methodologies to overcome some common challenges in characterizing well performance and analyzing well spatial and temporal relationships in terms of well spacing, stacking and infill timing. Ultimately, all the data is regularized to be ready for machine learning. The framework enables a rich set of state-of-the-art machine learning techniques. More importantly, the integration of machine learning with geology, reservoir and development data in a visual environment enables very intuitive and interactive testing, validation and interpretation, which provides valuable insight and confidence for development decision making.
The framework has been extensively employed in Permian Basin for important technical studies such as evaluation of new formation, optimization of well completion and spacing, and even PUD reserve booking compatible with SPEE recommended reliable technology. Field case studies clearly demonstrate the applicability and efficiency of the framework as well as the predictability and insights the machine learning techniques offer.