Machine-Learning-Based Solution Predicts Fluid Properties for Gas-Injection Data

Abstract

_ This article, written by JPT Technology Editor Chris Carpenter, contains highlights of paper SPE 211080, “Machine-Learning-Based Prediction of Pressure/Volume/Temperature Fluid Properties for Gas-Injection Laboratory Data,” by Kassem Ghorayeb, SPE, American University of Beirut and Schlumberger; Kristian Mogensen, SPE, ADNOC; and Nour El Droubi, Schlumberger, et al. The paper has not been peer reviewed. _ While machine learning (ML) is used extensively to predict black-oil properties, it is used less often for compositional reservoir properties, including those related to gas injection. Can typically extensive conventional laboratory data be used to help predict the necessary gas-injection parameters? This question is addressed in the complete paper. The authors present an ML-based solution that predicts pertinent gas-injection studies from known fluid properties such as fluid composition and black-oil properties—that is, learning from samples with gas-injection laboratory studies and predicting gas-injection fluid parameters for the remaining, much larger data set. Methodology The objective of the ML component is to predict the results of the swelling test using compositional and black-oil properties. Fig. 1 illustrates the three main steps that make up the work flow. The swelling-test data is processed and analyzed for use in the ML training as the first phase in the data-preparation and -analysis process. The following step is ML training, where features are engineered to examine correlations in the data, followed by the training of ML models using selected features and evaluation of the results. The swelling-test data for each fluid sample will be predicted using the trained ML models in the last stage, which creates a digital-twin database for swelling-test results. The three steps are detailed in the complete paper. Results and Discussion Several approaches were taken while training the ML model. One approach was to either cluster before the ML training, with one ML model trained for each cluster, or train a single ML model using all the data. The other approach was to use the chain method for the prediction of the outputs. In the chain approach, the outputs are given a specific order and, whenever one output is predicted, it is added to the inputs of the following output. The chain approach is particularly useful when the outputs are correlated.