| Abstrakt: |
Most available artificial intelligence models for bubble-point pressure (Pb) and oil formation volume factor (Bob) are multiple-input single-output models that are not reproducible or with complex topology. In this study, a multiple-input multiple-output genetically optimized neural network to predict oil Pb and Bob was developed using solution gas–oil ratio, API-oil gravity, gas gravity and reservoir temperature from 781 published datasets. The overall performance of the developed neural network resulted in a correlation coefficient (R) of 0.99918 and a mean square error (MSE) of 5.9941 × 10–4. The developed Pb and Bob models predictions resulted in a coefficient of determination (R2) of 0.9985 and R-value of 0.99423 for Pb, then R2 of 0.9842 and R-value of 0.99207 for Bob. Also, the Pb model had MSE of 2.0 × 10–4, root mean square error (RMSE) of 4.45 × 10–3 and average absolute percent relative error (AAPRE) of 0.4447, while Bob had 1.0 × 10–4, 3.9 × 10–3 and 0.0391 for MSE, RMSE and AAPRE, respectively. Furthermore, the generalization capacity of the developed models with new datasets resulted in R2, R, MSE, RMSE and AAPRE of 0.9613, 0.98046, 0.03713, 0.1927 and 19.2698, respectively, for Pb, as the Bob model had R2 of 0.9499, R of 0.97463, MSE of 0.0218, RMSE of 0.14763 and AAPRE of 8.20211. Again, the developed Pb and Bob models' generality performance showed that they outperformed some published correlations: Standing [1], Glaso [2], Al-Marhoun [3], Dokla and Osman [4], Almehaideb [5], Al-Shammasi [6], etc. Thus, these developed explicit Pb and Bob models for oil reservoir PVT prediction can be applied. [ABSTRACT FROM AUTHOR] |
