Abstract

Current data-driven machine learning (ML) methodologies used for forecasting production streams in unconventional wells are typically based on tabular features such as proppant load and well depth and focused on single-well performance. With the industry move to large multiwell development programs, ML models have been developed to account for the relative spatial position of wells in the "wine-rack" cross-sectional (barrel) view. These models also depend on engineered features such as staggered offset and vertical offset. Our novel approach uses 2D convolutional neural networks to encode geometric aspects of the wine-rack geometry as opposed to engineered geometric features.

Introduction

Rock cube development is a technique that involves drilling multiple horizontal wells in different layers of an unconventional reservoir, such as shale or tight oil, and then fracturing them simultaneously to create a large, stimulated rock volume (SRV) (Jacobs 2019). This technique aims to maximize the recovery of hydrocarbons from low-permeability formations by increasing the contact area between the wellbore and the reservoir (Salama et al. 2017). The optimal well spacing and fracture design are still challenging issues that depend on the reservoir properties, fluid characteristics, and operational parameters (Zhi-dong Yang et al. 2020).

To solve this problem, we present a convolutional neural network (CNN) solution that focuses on analyzing geometric components associated with rock cube development.

Theory and/or Methods

The methods of this investigation are summarized as 1) performing spatial-temporal density base clustering to create the drilling units (also referred as rock cubes or tanks); 2) engineering the relevant features, such as geometry and stimulation, at each drilling unit; 3) creating a rasterized 2D matrix of the "wine-rack" geometry of each drilling unit; 4) embedding stimulation features in the matrix to create a multidimensional tensor compatible with a CNN workflow; 5) training a CNN model for regression using the wine-rack rasters as inputs and monthly production of each drilling unit as targets; and 6) evaluating optimization with real data and creating what-if forecasting scenarios.

This content is only available via PDF.
You can access this article if you purchase or spend a download.