Machine Learning Models for Real-Time Skin Factor Prediction in Foamed Acid Treatments

For matrix acidizing operations an appropriate acid implementation requires careful planning to minimize well damage while preventing acid overconsumption. Machine learning algorithms allow realtime stimulation result assessments to accomplish this objective. This research seeks to establish advanced machine learning models which provide precise real-time readings for skin factor together with bottom-hole pressure at the coiled tubing nozzle during foamed acid treatments. This research built nine advanced machine learning models which used standard matrix acidizing operational parameters. Real bottom-hole pressure measurements obtained from deployed pressure gauges formed the basis for training the developed models. The predicted bottom-hole pressure enabled the determination of real-time skin factor measurements. The development of machine learning algorithms took place through analysis of an extensive dataset obtained from 31 wells. Through comparison of predictive bottom-hole pressure models with actual field measurements the most efficient machine learning methods demonstrated exceptionally low root mean square error. The Neural Network, AdaBoost, Random Forest and K-Nearest Neighbor algorithms produced root mean square error values of 0.047, 0.048, 0.054 and 0.058 respectively. This study introduces novel methods of using advanced machine learning models as predictive tools to track bottom-hole pressure and skin factor throughout matrix acidizing procedures with foamed acid. These predictive models act as accurate and swift replacements for the traditional pre- and post-stimulation well testing, conventional empirical multiphase flow correlations as well as mechanistic models and unified models. The proposed models function as an economical replacement for downhole pressure gauges as well because they eliminate both cost and duration issues.