Machine Learning for Determining Remaining Oil Saturation Based On C/O Spectral Logging in Multilayer String Cased Well

Dynamic monitoring of reservoir can reflect the physical response of fluid underground and clarify the oil and water distribution of the production, which is important for the production management in the later stage of oil field. C/O logging plays an important role in downhole fluid dynamic monitoring, which can accurately determine the oil saturation of the formation. However, due to the limitations of neutron generator yield and measuring speed, the statistical fluctuation in low and medium pore formation are large, masking the difference of spectrum measured and C/O response in oil and water formations; meanwhile, complex borehole and string combinations as well as lithology cause high uncertainty of oil-bearing saturation interpretation by conventional C/O interpretation methods. To improve the signal-to-noise ratio of measurements and the accuracy of oil-bearing saturation interpretation under complex borehole and formation conditions, machine learning methods for spectrum noise reduction and formation saturation prediction of C/O logging are introduced in this paper. Based on numerical simulation methods and calibration well measurements, the standard spectrum and the corresponding noise-add spectrum are constructed under different string and formation conditions. The main components in the noise-add spectrum are extracted and recovered using stacked autoencoder networks to realize the noise reduction process. Utilizing the ensemble learning method, artificial neural network and random forest method are integrated to construct the oil saturation prediction model, and the window counts of C, O, Si, Ca and Fe and the count ratios of C/O, Si/Ca and Si/Fe are obtained from the noise reduction spectrum, which are used as input features together with the string, mineral content, formation porosity, shale content and other bare-hole well data for oil content prediction of the formation saturation prediction. More than 5000 simulated data were calculated covering different string, formation lithology, porosity, and oil saturation; meanwhile, Wells with different string and lithology conditions were used as machine learning training samples to jointly construct the saturation prediction model. The model was validated by establishing simulated data under new parameters of different bole hole and formation conditions. The results showed that the autoencoder method improved the signal-to-noise ratio (SNR) of spectrum compared with Savitzky-Golay method, and the absolute error of saturation prediction from simulations and wells in the formation with porosity at 10 % is less than 10 %. This study improves the accuracy of C/O logging in determining the oil saturation of a formation under complex borehole and formation conditions.