Hydraulic Fracturing Fluid Optimization Using Machine Learning Approach: A Case Study of Marcellus Shale

Abstract

Abstract Friction reducer (FR) is a chemical additive utilized in hydraulic fracturing operations to minimize friction between the fracturing fluid and the wellbore walls. Its purpose is to overcome tubular drag during high-flow-rate pumping in stimulation treatments. In recent years, there has been a growing preference in the oil and gas industry for high viscosity friction reducers (HVFRs) in fracturing fluids due to their operational and economic benefits. While slick water fracturing fluids contain low concentrations of friction reducers, the concentration of FR used in shale gas reservoirs typically ranges from 0.5 to 2.0 gallons per thousand gallons (gpt). However, the optimal FR concentration for each stage has not been extensively studied. As a result, many oil and gas companies tend to use higher FR concentrations than necessary as a precautionary measure to achieve the desired injection rate and prevent screen out. Unfortunately, this practice leads to excessive FR usage during stimulation treatments, resulting in significant economic losses. The primary objective of this study is to gain a comprehensive understanding of the performance of the Hydraulic Fracturing Friction Reducer used in the completion of six horizontal Marcellus Shale wells. To achieve this, the data collected from wells drilled and completed in Marcellus Shale Energy and Environment Laboratory (MSEEL) project (Boggess pad) with over 340 hydraulic fracturing stages will be evaluated using different Machine Learning (ML) techniques. This study was conducted in three phases, employing a range of machine learning techniques to develop the optimal model for predicting surface treating pressure during hydraulic fracturing and use that to optimize the FR concentration. In phase 1, utilizing data from stage 1 of Boggess 1H, XGBoost emerged as the best-performing model, demonstrating high accuracy in predicting surface treating pressure. In phase 2, when deploying the same model on all other stages of Boggess 1H, promising results were observed. However, phase 3 revealed that extending the model to other adjacent wells resulted in unsatisfactory results, highlighting the necessity for new and distinct models tailored to each individual well.