Operators producing oil and gas from unconventional reservoirs tend to use a geometric completion design, which includes a constant stage length, cluster spacing, and number of clusters throughout a horizontal wellbore. The number of resulting perforations at these clusters that produce hydrocarbons varies from 15%-70% of the total number of perforations along the wellbore. This phenomenon results in lost revenue due to decreased production, and the sunk cost of paying for unsuccessful completions. What is causing unsuccessful induced hydraulic fracture propagations at these perforated clusters, and can it be detected and optimized? This study develops a machine learning algorithm to study the impact of anisotropic and heterogenic nature of shale on unconventional completion designs.
Using the Machine Learning method C-Means Fuzzy Clustering, well logging data, including lithological logs and geomechanical logs, can be interpreted to identify areas of high shale anisotropy and heterogeneity. C-Means Clustering will group measured data to other ‘like’ sets of data along the wellbore to create a predefined number of classifications, or clusters. C-Means is a fuzzy clustering method, meaning a datapoint's membership to each class assignment will vary from 0 to 1, where each datapoint is not a full member to a class, rather a partial member to every class. By measuring the magnitude of change in classification for datapoint across the logged interval, the change in shale properties was quantified as the Shale Heterogeneity and Anisotropy Indicator Value (SHAIV).
The initial part of the study was to validate that the SHAIV was effective at identifying non-productive clusters. To do this, a case study was preformed using the publicly available logging data from a 5800 ft Marcellus wellbore. After clustering the logging data and calculating the SHAIV across the wellbore, the SHAIV was compared to the production log on a per cluster basis in order to identify a correlation between non-productive clusters and the SHAIV at those clusters. It was found that 88% of the clusters that produced zero gas occurred in areas of high SHAIV, indicating a strong correlation between the SHAIV and clusters that fail to propagate and produce gas. The second part of the study is to apply the trained C-Means model to optimize the cluster placements and stage spacings.
This study proposed a robust machine learning algorithm to quantitatively relate the changes in shale properties, even in small distance intervals, with the seemingly random failure to propagate fractures for geometrically designed perforated clusters. Furthermore, the C-Means model can be implemented on the target horizontal wellbore to optimize the cluster placements of existing stages, or design new stages based on the C-Means classifications and SHAIV. In conclusion, operators could increase the number of successful perforations by running logs on horizontal wells in order to use C-Means Fuzzy Clustering. The additional cost per well of running these logs could quickly be offset by the additional revenue gained from increasing production with an optimized per well completions design.