Big-data analytics for production-data classification using feature detection: Application to restimulation-candidate selection

In recent years, there has been a proliferation of massive subsurface data sets from sources such as instrumented wells. This places significant challenges on traditional production-data-analysis methods for extracting useful information in support of reservoir management and decision making. In addition, with increased exploration interest in unconventional-shale-gas reservoirs, there is a heightened need for improved techniques and technologies to enhance the understanding of induced- and natural-fracture characteristics in the subsurface, as well as their associated effects on fluid flow and well performance. These challenges have the potential to be addressed by developing big-data-analytics tools that focus on uncovering masked trends related to fracture properties from large volumes of subsurface data through the application of pattern-recognition techniques. We present a new framework for fast and robust production-data classification, which is adapted from a real-time face-detection algorithm. This is achieved by generalizing production data as vectorized 1D images with pixel values proportional to rate magnitudes. Using simulated shale-gas-production data, we train a cascade of boosted binary classification models that are capable of providing probabilistic predictions. We demonstrate the viability of this approach for identifying hydraulically fractured wells that have the potential to benefit from restimulation treatment. The results show significant improvements over existing type-curve-based approaches for recognizing favorable-candidate wells, using only gas-rate profiles.