Application of modified GMDH network for CO2-oil minimum miscibility pressure prediction

Abstract

ABSTRACTThe CO2-oil minimum miscibility pressure (MMP) is an important parameter in the process of miscibility phase oil displacement. For the purpose of getting a high precision and better approximate expression, the group method of data handling (GMDH) network has been utilized to predict CO2-oil MMP in both pure and impure CO2 injection conditions. The transmittal mode and screening mode in each layer for variables were optimized in the proposed network. Input parameters considered as effective variables on CO2-oil MMP included reservoir temperature, crude oil composition and injection gas composition. Training and testing performances of the modified GMDH network were carried out using normalized parameters that collected from the literatures. The predicting outcomes were compared with those obtained using traditional equations, gene expression programming (GEP), back propagation neural network (BPNN) and traditional GMDH network. The modified GMDH network predicted the CO2-oil MMP with lowest error (pure CO2 condition: RMSE = 1.636, MAPE = 0.136% and SI = 0.094; impure CO2 condition: RMSE = 1.26, MAPE = 0.1378%, and SI = 0.079) and higher accuracy (pure CO2 condition: R = 0.971; impure CO2 condition: R = 0.973) than those predicted using equations and traditional neural networks (GEP, BPNN and GMDH).