Application of artificial neural network models and random forest algorithm for estimation of fracture intensity from petrophysical data

Abstract

Natural fractures play an essential role in the characterization and modeling of hydrocarbon reservoirs. Modeling fractured reservoirs requires an understanding of fracture characteristics. Fractured zones can be detected by using seismic data, petrophysical logs, well tests, drilling mud loss history and core description. In this study, the feed-forward neural networks (FFNN), cascade feed forward neural networks (CFFN) and random forests (RF) were used to determine fracture density from petrophysical logs. The model performance was assessed using statistical measures including the root mean squared error (RMSE), coefficient of determination (R2), mean absolute error (MAE), Kling Gupta efficiency (KGE) and Willmott’s index (WI). Conventional good logs and full-bore micro-resistivity imaging data were available from three drilled wells of the Mozduran reservoir, Khangiran gas field. According to the findings of this research, the FFNN model showed a higher KGE and WI, and a higher correlation coefficient (R2) compared to the CFNN model. The CFNN model outperformed the FFNN model with lower neurons. The models' performance was also improved by increasing the number of neurons in the hidden layers from 8 to 35. The findings of this study demonstrate that the measured and FFNN calculated fracture intensity is in excellent agreement with image log results showing a correlation coefficient of 92%. The RF algorithm showed higher stability and robustness in predicting fracture intensity with a correlation coefficient of 93%. The results of this study can successfully be used as an aid in a more successful reservoir dynamic modeling and production data analysis.