While diagnostic fracture injection tests (DFIT) data is relatively rare, pressure and flow rate data are widely collected for hydraulic fracture treatment stages in multiple transverse-fracture horizontal wells, both during pumping and during pressure falloff after the end of pumping. Recent papers have shown value in applying analysis developed for DFIT data to hydraulic fracture treatment falloff (HFTF) data. As is often the case with routine processing of large volumes of data, the time required for analyzing HFTF datasets for each of the treatment stages in a long horizontal well may be overly expensive in time and human resources. To address this issue, this paper applies machine learning techniques to standard HFTF pressure data, to produce data-driven mathematical models that estimate wellbore and perforation friction loss, near-wellbore tortuosity friction loss, and the instantaneous shut-in pressure (ISIP). The models are trained based on a small subset of hydraulic fracture treatment stages, data from for which is processed manually, to provide targets for model outputs. Model structure is based on multivariate statistical methods, including principal component regression (PCR) and partial least squares regression (PLSR).
Comparison between state-by-stage friction losses and ISIP values estimated manually and those quantified by machine learning reveals good agreement and underscores the value of the machine learning approach as a practical tool for field application tasks, including well and pad completion design decisions related to perforating strategies, and fracture and well spacing.