Predicting formation damage of oil fields due to mineral scaling during water-flooding operations: Gradient boosting decision tree and cascade-forward back-propagation network

Abstract

Water-flooding is one of the main options employed by the oil industry to meet the world's ever-increasing demand for oil, as the primary source of energy. This approach is highly prone to cause formation damage if the injected water is not compatible with the formation brine. In this study, decision tree optimized with gradient boosting (GBDT), cascade-forward back-propagation network (CFBPN), and generalized regression neural networks (GRNN) were employed, as relatively novel intelligent models, for the first time to develop accurate models to estimate the formation damage during a waterflooding operation in terms of damaged permeability. To compare the performance of these models, radial basis function (RBF) and multilayer perceptron (MLP) neural networks were also developed. The Levenberg-Marquardt algorithm (LMA), scaled conjugate gradient (SCG), and Bayesian regularization (BR) were used for training the MLP and CFBPN models. The results of this study showed the outperformance of the proposed GBDT model compared to the other developed models as well as previously proposed ones with an average absolute percent relative error (AAPRE) of 0.1465 % and correlation coefficient (R 2 ) of 0.9991 for the whole dataset. According to the results, the accuracy of the developed models could be ranked as follows: GBDT > CFBPN-LM > CFBPN-BR > RBF > MLP-LM > GRNN > MLP-BR > CFBPN-SCG > MLP-SCG. Moreover, it was shown that the GBDT mode could estimate more than 90 % of points with an absolute relative error of lower than 0.5 %. The trend analysis showed the high capability of the developed models in detecting the physical trend of the formation damage with variation of inputs. Then, the variable impact analysis was performed for this model, and the results reflect the high dependency of the model's predictions on the volume of injected water (Vinj), initial permeability (Ki), and ionic concentration of sulfate. Lastly, the Leverage approach was employed to determine suspected points as well as the applicability realm of the GBDT model. The results of the outlier detection indicated that only 4 points (0.93 % of the dataset) were detected as outliers, and the applicability realm of the proposed GBDT was verified. The findings of this communication shed light on the application of intelligent models and their power in predicting the formation damage caused during water-flooding operations before their occurrence. • Formation damage during water-flooding is modelled using GBDT, RBF, MLP, CFBPN, and GRNN. • LMA, SCG, and BR are used for optimizing MLP and CFBPN. • GBDT outperforms all models with AAPRE of 0.14 %. • The Leverage approach was used to determine suspected points and as the applicability realm of the GBDT model.