Data-Driven Prediction Method of Water Cut Based on Random Forest Regression Model

Abstract

Water cut (WCT) is a key parameter to analyse the performance of wells and reservoirs within a producing oilfield. However, the WCT data recorded in the life term of a well may not always be accurate or available, which may lead to the potential problem with well and reservoir models constructed with the data. This can lead to errors in predicted future well and field production, or missed opportunities for well workover activities. This paper describes a case study where the WCT of producing oil wells from a large Middle Eastern oil reservoir was modeled using random forest regression in order to identify errors and improvements in the field data. Pressure data and fluid properties were input as training variables and the model was evaluated by cross-validation. The relative importance of these variables was calculated and the coefficient of determination (R2) between the observed and predicted WCT of the test set was used to evaluate the model performance. It was found that the apparent density of the producing fluid and the variables related to the fluid composition have strong connections with WCT, as would be expected based on traditional vertical pipe flow theory. For the wells with good field WCT data the model accurately matched the real field data. For the wells with poor or absent field WCT data the model was used to predict the WCT and significantly enhance the dataset with a high degree of confidence. It is concluded that the random forest regression model can predict the WCT based on other well surveillance data. Overall, the current study provides an approach to integrate multiple factors of surveillance data to calibrate the WCT data, and can add significant value to well and reservoir models for the purpose of accurate production dynamics analysis and forecasting.