Comparative Analysis of Intelligent Algorithms to Predict the Minimum Miscibility Pressure for Hydrocarbon Gas Flooding

Abstract

Abstract Over the past two decades, enhanced oil recovery (EOR) has become one of the prominent techniques to increase the recovery of reservoir. EOR consist of four different techniques i.e. thermal, chemical, microbial and gas injection. Gas injection/flooding is one of the most robust applied technique for light oil EOR. Gas injection/flooding comprise of two processes called miscible gas flooding and immiscible gas flooding. Miscible gas flooding is the process in which both injection fluid and reservoir fluid are miscible. The minimum pressure at which both injecting, and reservoir fluids are miscible–is commonly known as Minimum Miscibility Pressure (MMP). The estimation of MMP is the challenging and crucial task in the designing of miscible gas flooding. In this study, we used experimental data along with the machine learning algorithms to find out the relation for MMP. Moreover, the comparison between three different algorithms (Support Vector Machine (SVM), Functional Network (FN) and Artificial Neural Networks (ANN)) was performed based on the results of statistical analysis. A new empirical correlation was established to estimate MMP as a function of reservoir temperature, reservoir oil composition, and injected gas composition. Since the data set contains reservoir composition data, the developed correlation incorporates the condensing/vaporizing mechanism during the miscible gas flooding process. The data set used to establish the new empirical correlations was based on experimental data obtained from literature. The data set was separated into two parts called development data and testing and validation data. To establish the new correlations, development data comprising of 70% of the data was used. Whereas, the rest 30% was kept solely to perform testing and validation of the developed correlations. Three different machine learning algorithms called Artificial Neural Networks (ANN), Support Vector Machine (SVM), Functional Network (FN) were used to develop the new correlation. The parameters of each algorithm were optimized to find out the best correlation. For ANN, the number of neurons, weights, and bias were optimized. Whereas for SVM, the epsilon and kernel parameters were tweaked to yield an accurate model. Likewise, for FN model a backward elimination method was found to be the best learning algorithm. To assess the performance of the developed correlations, statistical analysis was performed. Moreover, to avoid the occurrence of local minimum, multiple realizations (total 5000) with different algorithm parameters were run. The results indicated a minimal and acceptable average absolute error. Based on error, ANN was found to give the best correlation for prediction of MMP.