Machine learning–driven prediction of profile control effect in oil reservoir based on relative production difference

Traditional methods for predicting profile control effect have limited accuracy, with deviations of varying degrees between predicted and actual oil production after operation. In order to solve this problem, this study introduces a prediction method of profile control effect based on relative production difference (PRD). Initially, Long Short-Term Memory (LSTM) network, Gated Recurrent Unit (GRU) network, and Transformer neural network were used to predict the benchmark production of water flooding. Subsequently, the study identified factors influencing profile control effect as input features and the relative difference between benchmark production of water flooding and post-treatment actual production as the target output. Then three machine learning models were constructed for prediction of profile control effect, namely Adaptive Boosting (AdaBoost), Extreme Gradient Boosting (XGBoost), and Random Forest (RF). Hyperparameters of all models were optimized using the Optuna framework during training process. Finally, the study applied the optimal profile control effect prediction model to predict the profile control effect in S reservoir. The results show that LSTM performs the best in predicting the benchmark production of water flooding, with R² values of 0.9874 and 0.9508 for the training and test sets, respectively. RF performs the best in predicting the profile control effect, with R² values of 0.9988 and 0.9995 for the test and cross validation sets, respectively. In oilfield applications, the test set R² of RF exceeds 0.9, which proves that the predicted production after profile control is highly consistent with the actual production, and the model has strong generalization ability.