Effects of Data Preprocessing on the Prediction Accuracy of Lubricant Service Life of Water Injection Pump in Enhanced Crude Oil Recovery Facility

Data mining approaches are widely adopted in various fields, including in the petroleum industries. One of the significant problems in the enhanced crude oil recovery facility is reliability of the water injection pumps. Lubricant service life has important role in this reliability issues. This paper reports the effects of the data preprocessing and the regressor on the predictions of lubricant service life of water injection pump in enhanced crude oil recovery facility. As results, the performance of the algorithm is the decision tree regressor with R
²
= 0.88, then the random forest regressor with R
²
= 0.82, and the gradient boosting regressor with R
²
= 0.71. The lubricant data for the DE (drive end) pump section contains more noise than the NDE (non-drive end) pump section. This indicates that the difference in the bearing components of the pump NDE compared to other parts that are lubricated does not significantly affect the characteristics of the lubricant service life. For small amounts of data, a simpler algorithm such as decision tree can produce better performance than ensemble algorithms like random forest and gradient boosting. Ensemble method such as random forest can outperform the decision tree for data that contains a lot of noise such as lubricant data on the pump. Moreover, the parameter in the lubricant that are most related to the service life of the lubricant on the motor are Viscosity100, Zinc, and Copper. While the content of the lubricant that is most related to the service life of the pump is Viscosity100, Zinc, Boron, and Viscosity40.