Hidden Markov approach for screenout detection in unconventional reservoirs, A

Abstract

Hydraulic fracturing has gained its popularity all over the world as more tight geologic formations are developed economically for hydrocarbon resources. While exploring for new unconventional resources such as shale plays, horizontal drilling and multi-stage hydraulic fracturing are required to stimulate the geologic units to increase well production. However, due to the stages' operating complexity, different kinds of disruptions in fracturing operations may occur and even result in great economic loss. Screenout is one of the issues caused by the blockage of proppant inside the fractures. In this project, a horizontal well landing in the Niobrara B shale, Denver-Julesburg (DJ) Basin, is simulated with multiple fracturing stages in a hydraulic fracturing software and various synthetic fracturing treatment data are forward modeled for both screenout and non-screenout scenarios. This thesis describes a screenout classification system based on Gaussian Hidden Markov Models, trained on simulated data, in order to predict screenouts and provide early warning by learning pre-screenout patterns in the simulated surface pressure signals. The classification system consists of two Gaussian Hidden Markov Models (screenout and non-screenout), each of which is fitted and optimized by its respective training set. Both Hidden Markov Models are assigned with two 1D Gaussian probability density functions to represent the distribution of their associated simulated surface pressure signals. During the classification process, once a new surface pressure sequence is observed, the log likelihood is calculated under both fitted models and the model with greater likelihood will be predicted as the class of this new observation. The classification system is validated and verified with a hold-out testing data set from the simulations and the statistics of the performance is presented in a confusion matrix. The results indicate the classification system achieves 86% accuracy for successfully predicting screenout events around 8.5 minutes prior to screenout occurring in the simulation. The described methodology is demonstrated to be a useful tool for early screenout detections and shows its promising feasibility of other time-series analysis such as microseismic data.