--> Building Reliable Local Earth Models of Utica Shale Reservoirs in the Presence of Missing Values and Sparse Data: An Integration of Classic Statistics, Spatial Statistics, and Machine Learning

2019 AAPG Eastern Section Meeting:
Energy from the Heartland

Datapages, Inc.Print this page

Building Reliable Local Earth Models of Utica Shale Reservoirs in the Presence of Missing Values and Sparse Data: An Integration of Classic Statistics, Spatial Statistics, and Machine Learning

Abstract

Well placement and completion engineering are keys to successful drilling and economic development in the Oil and Gas business. Subsurface earth models of reservoirs have traditionally been the collaboration point for geoscientists and engineers in conventional plays. However, such models depend on “enough” data from well logs and seismic profiles to build a digital model that depicts subsurface measurements with manageable uncertainty. In unconventional reservoirs, however, these data sources are often lacking, particularly at local project scales. The Utica Shale, an unconventional “source-reservoir” in the Appalachian Basin, is a good example to demonstrate the principle problems facing subsurface modeling of such reservoirs, specifically, sparse data (missing log and seismic data), and missing values (unaccounted-for missing values within the data present). Although similar, these problems can have quite different impacts on modeling, although both may result in increased model uncertainty, particularly as the industry aggressively pushes towards machine learning (ML) and automation. Using the Utica Shale in eastern Ohio, we first focus on the management of missing values, followed by subsurface modeling of Utica formation including the Point Pleasant. Missing values are handled largely by classical statistical approaches using listwise-pairwise deletion of data to compile a reduced but data-complete set of properties. When placed in a ML context, the goal is to limit the negative impact of data deletion. We present an evaluation of the classification of missing value types (causes) within a Utica Shale project along with a variety of ML appropriate statistical methods for imputing missing values. When modeling the subsurface, well log data is the principle source of information, along with seismic data. In unconventional resource evaluation, these data are often unavailable due to the expense of acquiring well log data in horizontal wells, resulting in sparse data. This results in a data pool of some historical well logs and a paucity of horizontal logs. In such cases, building meaningful subsurface models is nearly impossible. Here, we demonstrate the utility of integrating a local 3D Burial History Model and its derived finite volume properties along with geostatistical co-simulation of local available data to produce reliable models of the subsurface and associated uncertainty for well placement and engineering collaboration.