Summary
Sand screenout is one of the most serious and frequent challenges that threaten the efficiency and safety of hydraulic fracturing. Current low prices of oil/gas drive operators to control costs by using lower viscosity and lesser volumes of fluid for proppant injection—thus reducing the sand-carrying capacity in the treatment and increasing the risk of screenout. Current analyses predict screenout as isolated incidents based on the interpretation of pressure or proppant accumulation. We propose a method for continuous evaluation and prediction of screenout by combining data-driven methods with field measurements recovered during shale gas fracturing. The screenout probability is updated, redefined, and used to label the original data. Three determining elements of screenout are proposed, based on which four indicators are generated for training a deep learning model [gated recurrent units (GRU), tuned by the grid search and walk-forward validation]. Training field records following screenout are manually trimmed to force the machine learning algorithm to focus on the prescreenout data, which then improves the prediction of the continuous probability of screenout. The Pearson coefficients are analyzed in the STATA software to remove obfuscating parameters from the model inputs. The extracted indicators are optimized, via a forward selection strategy, by their contributions to the prediction according to the confusion matrix and root mean squared error (RMSE). By optimizing the inputs, the probability of screenout is accurately predicted in the testing cases, as well as the precursory predictors, recovered from the probability evolution prior to screenout. The effect of pump rate on screenout probability is analyzed, defining a U-shaped correlation and suggesting a safest-fracturing pump rate (SFPR) under both low- and high-stress conditions. The probability of screenout and the SFPR, together, allow continuous monitoring in real time during fracturing operations and the provision of appropriate screenout mitigation strategies.
Evaluation of brittleness index (BI) is a fundamental principle of a hydraulic fracturing design. A wide variety of BI calculations often baffle field engineers. The traditional value comparison may also not make the best of BI. Moreover, it is often mixed up with the fracability in field applications, thus causing concerns. We, therefore, redefine fracability as the fracturing pressure under certain rock mechanical (mainly brittleness), geological and injecting conditions to clarify the confusion. Then, we propose a data-driven workflow to optimize BIs by controlling the geological and injecting conditions. The machine learning (ML) workflow is employed to predict the fracability (fracturing pressure) based on field measurement. Three representative ML algorithms are applied to average the prediction, aiming to restrict the interference of algorithm performances. The contribution of brittleness on pressure/fracability prediction by error analysis (rather than the traditional method of BI-value comparison) is proposed as the new criterion for optimization. Six classic BI correlations (mineral-, logging- and elastic-based) are evaluated, three of which are optimized for the derivation of a new BI using the backward elimination strategy. The stress ratio (ratio of minimum and maximum horizontal principal stress), representing the geological feature, is introduced into the derived calculation based on the independent variable analysis. The reliability of the new BI is verified by error analyses using data of eight fracturing stages from seven different wells. Approximately 40%~50% of the errors are reduced based on the new BI. The differences among the performances of algorithms are also significantly restrained. The new brittleness index provides a more reliable option for evaluating the brittleness and fracability of the fracturing formation. The machine learning workflow also proposes a promising application scenario of the BI for hydraulic fracturing, which makes more efficient and broader usages of the BI compared with the traditional value comparison.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.