Machine-Learning Assisted Analysis of Frac Water Hammer

Abstract

Abstract The purpose of this work was to evaluate if frac water hammer signature characteristics are representative of important hydraulic fracture and reservoir properties in a horizontal well and if those characteristics can be used as inputs in a predictive Machine Learning model. Water hammer is an oscillatory pressure signal generated as a result of an abrupt change in wellbore fluid velocity, for example at the end of a fracturing treatment when the pump rate is quickly dropped to zero. Authors developed an algorithm in python to automatically identify surface pressure and fluid pump rate channels in the raw data, detect and flag end of pumping events, parse out the water hammer and pressure decline signal from raw data. Numerical optimization algorithm was then used to approximate water hammer characteristics from the pressure signal based on the modified damped sine wave equation. The derived equation coefficients were used as inputs (features) in a Random Forest classification model to classify individual fracture stage contributions to a horizontal well's production profile. Production log tool (PLT) results from 8 horizontal unconventional wells and corresponding 1-second field data for 78 hydraulic stage fracture treatments were used in the study. The water hammer characteristics and parameters of the theoretical vs. actual curve match for each stage, such as initial amplitude of the signal, decay rate, phase angle, angular frequency, number of peaks, etc. were used as features for modeling. The data was split into 70 % −30 % train and test sets. A Random Forest Classifier model was trained on the train set to classify individual fracture stages in a horizontal well as either contributing or non-contributing to production. The model was validated against the test set with overall test classification accuracy of 0.71 and F1-Score of 0.72. Based on the study the authors conclude that water hammer characteristics derived from surface pressure signal via curve-matching technique can be useful for classification of fracture stage contribution to production in a horizontal well.