Accurate Prediction of CO2 Minimum Miscibility Pressure Using Adaptive Neuro-Fuzzy Inference Systems

Abstract In CO2 gas injection, the oil recovery rate of the miscible flooding is significantly higher than that of the non-miscible flooding. The miscibility of oil and CO2 can only be achieved when pressure is above the minimum miscible pressure (MMP), hence MMP is an important parameter for the optimal design of the CO2 injection in the reservoir. The MMP can be determined by traditional methods such as experimental and empirical correlation approaches. The experimental method is accurate but time-consuming and expensive, while the correlation methods are efficient but have limited application range due to its theoretical assumption's limitation. Thus, a more efficient and accurate method for MMP measurement is in need. Artificial neuro network (ANN) is an effective tool for engineering estimation, hence some MMP prediction model based on ANN was proposed by researchers, but the accuracy of models still can be improved. In this study, four factors (heavy hydrocarbon molecular weight, reservoir temperature, volatile and intermediate ratio oil) of MMP were designed as input, MMP as output, and ANFIS model is constructed for MMP prediction, the accuracy is estimated by the root mean square error (RMSE). The simulation results indicate that the ANFIS model has higher accuracy (average RMSE=1.846) and wider application range than those traditional correlation approaches (best-performed correlation RMSE=4.25). ANFIS runs faster and cheaper than the experimental approach. Among all the ANFIS models, the hybrid algorithm optimized ANFIS with Gaussian member function is most accurate with RMSE=1.44. BP algorithm optimized ANFIS with PSigmoid membership function produces the largest error (RMSE=2.83). Therefore, the ANFIS model developed in this study is more accurate and time-saving than traditional methods in predicting MMP and is more accurate than the previous neural network MMP prediction models.