Scale-Prediction/Inhibition Design Using Machine-Learning Techniques and Probabilistic Approach

Abstract

Summary This paper presents a data-driven methodology to predict calcium carbonate (CaCO3)-scale formation and design its inhibition program in petroleum wells. The proposed methodology integrates and adds to the existing principles of production surveillance, chemistry, machine learning (ML), and probability theory in a comprehensive decision workflow to achieve its purpose. The proposed model was applied on a large and representative field sample to verify its results. The method starts by collecting data such as ionic composition, pH, sample-collection/inspection dates, and scale-formation event. Then, collected data are classified or grouped according to production conditions. Calculation of chemical-scale indices is then made using techniques such as water-saturation level, Langelier saturation index (LSI), Ryznar saturation index (RSI), and Puckorius scaling index (PSI). The ML part of the method starts by dividing the data into training and test sets (80 and 20%, respectively). Classification models such as support-vector machine (SVM), K-nearest neighbors (KNN), gradient boosting, gradient-boosting classifier, and decision-tree classifier are all applied on collected data. Prediction results are then classified into a confusion matrix to be used as inputs for the probabilistic inhibition-design model. Finally, a functional-network (FN) tool is used to predict the formation of scale. The scale-inhibition program design uses a probabilistic model that quantifies the uncertainty associated with each ML method. The scale-prediction capability compared with actual inspection is presented into probability equations that are used in the cost model. The expected financial impact associated with applying any of the ML methods is obtained from defining costs for scale removal and scale inhibition. These costs are factored into the probability equations in a manner that presents incurred costs and saved or avoided expenses expected from field application of any given ML model. The forecasted cost model is built on a base-case method (i.e., current situation) to be used as a benchmark and foundation for the new scale-inhibition program. As will be presented in the paper, the results of applying the preceding techniques resulted in a scale-prediction accuracy of 95% and realized threefold cost-savings figures compared with existing programs.