Data-Driven Model for Measuring Hydraulic Fracturing Efficiency by Utilizing the Real-Time Treatment Data

Abstract

The paper presents a practical tool for hydraulic fracturing efficiency evaluation. The tool is based on a data-driven approach that helps in interpreting real-time data. Based on the hydraulic fracturing (HF) job monitoring, statistic metrics and key performance indicators (KPIs) are generated to be valuable input for further designs and identification of potential savings in operation. Machine learning (ML) algorithms are proposed to reduce the tedious work of completion engineers by automatically classifying each treatment schedule's timestamp and assigning the stage label. For operation stages classification Support vector machines and neural networks algorithms are used. These models are trained and evaluated on real-time treatment datasets. After automatic stage recognition, relevant statistic parameters are calculated, enabling advanced data analytics. Detailed analysis of historical data allows to identify the areas for improvements and set new best practices. The first research objective was to gather data from various companies and structure them under the same template to conserve the most critical information gained during the hydraulic fracturing job. Afterwards, the data are preprocessed and labelled by using signal processing routines that significantly decrease the labelling time. The labels or classes are used to define different stages that can be distinguished during the treatment. Finally, the goal is to decrease the necessary time for data labelling. Therefore, two multiclass classification models (Support Vector Machines (SVM) and Neural Network (NN)) are built and evaluated. Based on evaluation metrics, both models resulted in high accuracy and reliable results. However, the SVM model resulted in slightly higher accuracy and an F1 score. The key value of these models is that they provide a computational method to extract a pumping schedule from hydraulic fracturing time-series data automatically. Also, these models allow conducting post-job analysis and choosing the proper pump schedule for a future HF treatment based on previous experience. This past-job analysis could contribute to the effectiveness of future operations by utilizing the materials and fluids more efficiently.