A Practical Stacking Learning Framework with Hybrid Resampling for New-Well Artificial Lift Mode Selection

Summary Due to the long-term exploitation of oil and gas resources, the reservoir pressure will inevitably reduce to a level where it can no longer sustain acceptable production rates. Artificial lift provides additional energy to lift resources produced downhole to the wellhead. However, as the result of artificial lift selection will directly affect the productivity, it is crucial to select a suitable lift method. Generally, the more the parameters there are available for artificial lift selection, the higher the quality of the results will be. However, two challenges arise when selecting artificial lift methods for new wells: (A) Fewer parameters are available from new wells and (B) historical data tends to be multiclass imbalanced. To address these problems, we propose a hybrid resampling stacking ensemble learning (HRSE) framework and implement an intelligent optimization method for new-well artificial lift selection in the oil-recovery process. Specifically, HRSE mitigates the problem of limited available parameters by using stacking ensemble learning to extract valuable information from the data. Additionally, HRSE improves the accuracy of minority class samples through hybrid resampling techniques. To evaluate the performance of the proposed method, two imbalanced data sets from Daqing Oilfield were used, namely, Data Set A and Data Set B. Experimental results show that the HRSE framework achieves greater than 95% and 98% accuracy in the two minority classes of Data Set A, and with an overall average accuracy of 96%. For Data Set B, the accuracy across five minority classes reaches 95%, 91%, 97%, 94%, and 92%, respectively, and with an overall accuracy of 94%. Obviously, the proposed HRSE framework provides a more objective and scientific decision-making reference for petroleum engineers selecting artificial lift methods for new wells.