Predictive model for water absorption in sublayers using a machine learning method

Abstract

The sedimentary rhythm of Chinese oilfields is complicated and the heterogeneity is extremely strong. Allocating water absorption of each sublayer by dividing coefficient or numeric simulation cannot accurately reflect the actual water injection of the reservoir. Calculation based on water absorption profile monitored on site is the most commonly used method in oil field. However, access to these type of data is limited due to its cost and time related to acquisition. In this study, a machine learning approach was adopted to predict water absorption in sublayer based on geologic and production parameters of injectors and producers. On the one hand, it can save test costs. On the other hand, it can continuously predict water absorption of sublayers, and make up for water injection wells with insufficient injection profiles. A handful of training observations are obtained from on-site monitoring. Interwell connectivity is first conducted to identify connected producers for injectors. Introducing interwell connectivity helps to constitute predictor variables and yield significant improvements in feature selection. Connectivities in the well group are represented by similarity between injection sequence and production sequence, which is computed by Dynamic Time Warping. Average importance of predictors are then measured based on Mean Decrease Impurity, Mean Decrease Accuracy, and Ridge regression. Some relative important features are selected to consist the final predictors. The Extreme Gradient Boosting model is developed and then trained for making predictions given any set of observations. The proposed approach is validated by using actual field case from SL oilfield, China. Results show a significant correlation between predictions and actual value from on-site monitoring.