(468c) Decision Making for Unconventional Natural Gas Production: A Multivariate Analysis Approach

Recovering natural gas from unconventional resources, such as shales, tight gas, has recently increased the recoverable reserves in North America by a significant factor.  Made possible through reservoir stimulation by massive hydraulic fracturing, the development of such resources poses engineering and environmental challenges that must be addressed.  One such challenge is associated with making decisions, in a systematic way, about where to place and how to complete new production wells.  The current practice relies on using a variety of available data, from early seismic studies and core analysis, to production from existing wells over a period of time.  The entire process of data analysis requires significant human input, and can be quite laborious, with the possibility of not making optimal use of all available information.  To address this need, we have undertaken a collaborative project, aiming to develop a suite of integrated software modules that facilitate the integrated use of all available data for systematic decision making.  In this presentation we discuss our latest work on using multivariate data analysis and modeling methods that help the user analyze data from past production, so that factors affecting performance can be identified and subsequent decisions can be made to enforce factors that boost production while avoiding problematic ones.

The multivariate methods applied in this study are principal component analysis / partial least squares (PCA/PLS), Linear discriminant analysis (LDA), and symbolic regression (SR).  The analysis was performed on real production data from the Holly Branch Field in Freestone County, Texas.

Production data from twelve existing wells in the field were fed to PCA pool. One big cluster of similar performance wells and three distinct behavior wells were obtained as a result of this analysis. Probing further into well completion data clearly demonstrated the reason for this distinction of the three wells from the rest of the nine-well cluster.  It was also found that production data from all wells follow the stretched-exponential power-law decline curve (q(t)=q_i (exp(-D_i tⁿ)) ) . Once these decline curve parameters were estimated, correlations and models were developed using PLS, LDA and SR with decline curve parameters q_i, D_i, and n as outputs and well properties such as porosity, permeability, fracture conductivity, fracture half-length, aspect ratio, drainage area, and cumulative gas in place in drainage area, as inputs. Adjusted R² of more than 90% for each of the decline curve parameters was obtained which proves good predictive ability of the models. Thus, these models obtained can be used confidently to predict production behavior of new wells in this field if other well planning parameters are known. Finally, the estimated ultimate recovery (EUR) can be evaluated once the production decline parameters are known.

Future extensions of this work include additional testing using real data, and exploration of other multivariate methods that might help identify patterns in available production data.