Ensemble Learning with Feature Fusion for Well-Overflow Detection

Traditional methods for kick prediction rely on experience and basic physical models, but they struggle to maintain high accuracy under complex well conditions and varying geological environments. Moreover, conventional approaches often fail to fully leverage big data resources. Therefore, this study aims to improve the accuracy of kick prediction through machine learning models, especially by adopting an innovative feature fusion strategy to optimize the prediction model. We evaluated the comprehensive performance of six different machine learning models: GBM, LightGBM, CatBoost, XGBoost, Random Forest, and AdaBoost. Based on these evaluations, we proposed a feature score weighted fusion method. This method weights the importance scores of features based on the accuracy of each model, selecting the features that have the most significant impact on the prediction results. Experimental results show that the model utilizing the feature fusion strategy performs better on the test set than any single base model. Specifically, the GBM model, optimized through feature fusion, achieved better values in accuracy, F1 score, and F2 score, proving the effectiveness of this feature fusion strategy in enhancing the accuracy of kick predictions during well control operations. Furthermore, this study compared the performance of different models, providing valuable insights for kick prediction in well control operations. This research not only demonstrates the potential of feature fusion strategies in improving the accuracy of kick prediction but also offers new insights for the future application of machine learning technologies in the field of oilfield data analysis.