Prediction Method for Recovery Rate in Complex Fault-Block Reservoirs Based on Machine Learning Algorithms

Abstract Existing methods for predicting the recovery factor of complex fault-block reservoirs lack comprehensive consideration of influencing factors, leading to a higher recovery factor predicted by core flood experiments compared with the actual field recovery factor. To address this challenge, this study proposes a method for predicting the recovery factor of complex fault-block reservoirs based on machine learning algorithms, using the ZY oilfield in China as an example. This study first selects various factors affecting the recovery ratio of complex fault-block reservoirs from both development and geological perspectives. Next, the Spearman correlation analysis method is used to calculate the correlation coefficients of factors in order to identify the key controlling factors. Then, eight machine learning algorithm models are built based on the selected key factors, and four models with higher prediction accuracy are optimized using five evaluation metrics. Finally, the weighted average of the recovery factor predictions from the four optimized models is calculated using the coefficient of determination (R²), and a multiple linear regression method is used to establish a recovery factor prediction model for complex fault-block reservoirs. The results show that the key controlling factors affecting the recovery factor of complex fault-block reservoirs are: original oil saturation, permeability variation coefficient, well density, permeability, permeability ratio, and porosity, with the absolute values of their correlation coefficients being: 0.38, 0.30, 0.27, 0.25, 0.24, and 0.21, respectively. By comparing the five evaluation metrics (MSE, RMSE, MAE, MAPE, and R²), four models (CatBoost, XGBoost, DT, and GBDT) were selected, which showed better performance than other models, with R² values of 0.805, 0.856, 0.877, and 0.885, respectively. The recovery factor prediction model established in this paper was applied to 10 complex fault-block development units, along with six empirical formulas for water drive sandstone reservoirs. The error of the proposed model is only 8.50%, which is lower than the errors of the other empirical formulas, which are 14.14%, 23.93%, 22.89%, 15.45%, 48.41%, and 42.85%, respectively. It proves the accuracy and reliability of the established model in predicting the recovery ratio of complex fault-block reservoirs. The method provided by this paper serves as a basis for the prediction of recovery ratio in complex fault-block reservoirs, offering valuable guidance for oilfield development plan adjustments and well allocation.