Volume 26: Closing Carbon Cycles – A Transformation Process Involving Technology, Economy, and Society: Part I

Prediction of Minimum Miscibility Pressure (MMP) of CO2-Crude Oil Systems Considering the Differences of MMP in Different Experiments Based on Artificial Neural Network and Bayesian Optimization Algorithm Can Huang, Leng Tian, Lili Jiang, Wenxi Xu, Jiaxin Wang

Abstract

In this work, considering the differences of MMP in slim tube and RBA, an ANN-based technique has been developed to estimate MMP and identify the influence of different measurement methods to the predicted MMP for the compiled 193 sets of MMP data for CO2 and crude oil systems under various conditions. The 193 MMP datasets are collected from open literature, including 60 sets for slim tube and 134 sets for RBA. In addition to MMP, each group of the dataset mainly contains 12 influencing factors, which can be divided into the following five main categories, i.e., compositions of the injected gas (GCO2, GN2, GH2S, GCH4, and GHC), reservoir temperature (TR), molar fraction of each component in crude oil (LVOL, LINT, LC5-C6, and LC7+), molecular weight of C7+ oil components (MWC7+), and the method used to measure MMP (EM). The EM value of 0 indicates that the MMP is obtained through the slim tube experiment, and the MMP is measured by RBA when the value of EM is 1. To comprehensively improve the generalization ability of the model, the Bayesian optimization algorithm (BOA) was applied to optimize the model structure. Then, the developed ANN-BOA model was evaluated by comparing the prediction results with the measured MMPs and the predicted MMPs from the same model based on mixed MMP data, respectively. Compared to the existing model without taking the measurement method of MMP as input to generate the forecasting MMP data, the newly proposed model not only has the lowest overall MAPE of 6.84%, lowest overall MSE of 3.2062, and highest overall R2 of 0.9739 on the testing datasets for the three random runs, but also vividly reflect the interactive relationships of each influential factor and the MMP. Finally, the differences of MMP measured with the slim tube method and RBA method on the predicted MMP were analyzed. The results indicate that MMPs measured by RBA are generally higher than those measured by slim tube under the same reservoir conditions during pure and impure CO2 injection process, which explains why the prediction accuracy of the newly developed ANN-BOA model considering the influences of different measurement methods on MMP prediction is higher than that of the existing model.

Keywords Artificial neural network, minimum miscibility pressure, machine learning, CO2-Crude Oil System, bayesian optimization algorithm