(449e) Commercial Applications of Petroleum Data Analytics Using Shale Oil and Gas Data Sets from Permian Operator

This work was focused on developing novel methods to facilitate rapid data to knowledge transformation for subsurface applications, particularly those related to unconventional oil and gas. It was conducted at the U.S. Department of Energy’s (DOE) National Energy Technology Laboratory (NETL) in collaboration with industrial partner, a major shale-field operator in the Permian Basin, located in West Texas, United States. With progressing oil and gas field maturation and infill drilling, tighter well spacing to maximize resource utilization may lead to inter-well interference, to the point that induced fracture hits may start to affect the efficiency of fluid recovery. Fracture hit (aka, ‘frac hit’) is the form of hydraulic inter-well communication where hydrocarbons production at an existing well is affected by pumping hydraulic fracturing fluids into another well. Petroleum data analytics was used in this work to demonstrate the opportunities to optimize field development planning, including the frac-hit detection and analysis, support operational decision-making and improve resource recovery. It involved descriptive and predictive analytics which generated fuzzy patterns and key performance indicators for site characterization, completion design, hydrocarbons production, and frac-hit model implementation. Random Forest (RF) was used as the base estimator in the feature importance assessment, to evaluate the model input contributions to production response and to appraise data value. The demonstration also involved a combination of several Machine Learning (ML) algorithms such as Fuzzy Set Theory and Artificial Neural Networks (ANN). The ANN models included long short-term memory (LSTM), recurrent neural networks (RNN), and multilayer perceptron (MLP) neural networks to innovatively identify hydraulic communications between fracking child well(s) and producing parent well(s) in the same pad (intra-pad interactions) and/or in different pads (inter-pad interactions) based on pressure and production measurements in specified time series. The models took advantage of LSTM in handing temporal data and the MLP in feature extraction. The developed workflow can capture the time-variable features of frac hits when the model is deep and wide enough with enough trainable parameters for learning and feature extraction. The number of model parameters ranged from approximately 44,000 to 916,000. The five-fold increase in the number of model parameters (from the smallest to mid-size model) resulted in almost three orders of magnitude decrease in the mean-squared-error (MSE). The additional four-fold increase in the number of model parameters (from mid-size to the largest model) resulted in another order of magnitude decrease in MSE. As expected in time-series applications, temporal attributes such as the current-day lift-inlet pressure, daily lag pressure, and days online had orders of magnitude greater importance than static and categorical attributes. The results also demonstrated that trained ML models can accurately predict the frac-hit probability in near-real-time, with major implications for a potentially real-time operational support. The observed patterns and correlations provide a good starting point for determining the likely cause of observed shut-in an frac-hit events.