Applications of Machine Learning for Estimating the Stimulated Reservoir Volume (SRV)

Abstract

Hydraulic fracturing process is an integrated part of the wellbore completions in unconventional reservoirs. Typically, the process is designed before executing the job, aiming at optimizing the final fracture geometry and increasing stimulated reservoir volume (SVR). The physics-based models used for designing purposes are typically built on several simplified assumptions and do not match the SRV estimates from field observations. This work proposes a data-driven and machine learning-based approach for estimating SRV in unconventional reservoirs. A dataset from the Marcellus Shale Energy and Environment Laboratory (MSEEL) project is used in this study. The model’s input data include stimulation parameters of 58 stages of two wells (MPI-3H and MPI5H). The model output consists of the size of the corresponding microseismic (M.S.) cloud to each stage. Because of the limited number of stages (58) and to make the predictions close to near-real-time, each stage and its corresponding M.S. events are broken into steps, each having unique operational parameters (e.g., proppant and fluid volume and injection rate, among other parameters). This approach helped us to increase the number of required samples for data-based modeling to 829 samples. A standard procedure, including data cleaning, normalization, exploratory data analysis, and input data split (20/80), is then applied to the data. This is followed by various machine learning algorithms used to predict SRV. The respective performances of different methods are compared against each other. Microseismic (M.S.) monitoring is commonly used to monitor fracture topology evolution over time during the fracturing process. The recorded M.S. clouds that are observed during the hydraulic fracturing process can give a rough estimate of the stimulated reservoir volume (SRV). In this approach, a volume (or area in 2D) that encloses most of the M.S. events can be estimated at different time windows and used as the model output. All models were validated on the test set, and a good match was obtained. Our approach will be the first step toward real-time data-based modeling of “Dynamic SRV” or DSRV, which can be used to provide a better understanding of fracture propagation in unconventional reservoirs. It also can be used to optimize the well stimulation process before executing the job. Moreover, the developed model can be trained and used for other unconventional reservoirs.