Feature engineering process on well log data for machine learning-based SAGD performance prediction

The conventional numerical simulation requires a reliable three-dimensional reservoir model to evaluate the steam-assisted gravity drainage (SAGD) performance, but it may be time-consuming and need a significant amount of static data. This study aims to develop a machine learning (ML) model that predicts the SAGD performance based on well log data from a near single well. Although the ML model is useful in capturing underlying non-linear relationships between input features and target features, it requires a large number of training data. However, the number of data is usually small in commercial fields, limiting the availability of all log responses as input features to consider reservoir heterogeneity. In case of using small dataset, the overfitting can occur with an excessive number of model parameters. Therefore, this study introduces a feature engineering process to extract and select key features from well log data to alleviate the overfitting problem. In this study, the feature engineering process was optimized by sensitivity analysis on feature extraction and feature selection methods. For data preparation, the 63 SAGD well pairs and their matched logged wells were screened from the six commercial oil sand fields in Alberta, Canada. The SAGD process mechanism and petrophysics were considered to build data screening criteria. Each depth of the log data was categorized into payzone or shale barrier. For feature extraction, sampling intervals were placed on log data. This calculates the net pay thickness, averaged gamma ray, deep resistivity, and neutron porosity logs in the pay-zone. In addition, the elevation of two shale barriers and the distance between the SAGD well and logged well are extracted. Sensitivity analysis of the sampling interval length was conducted to determine the optimal conditions considering heterogeneity and feature reduction. For the feature selection, various combinations of the extracted features were used to examine their inherent meaning for predicting the SAGD performance. The multi-layer perceptron model was trained based on various selected features to predict the SAGD performance indicators of a seven-year operation. The optimal sampling interval length was determined as 10m and the combination of gamma ray, deep resistivity, shale barrier elevation, and the distance information was selected as key input features. The model showed reasonable prediction results for SAGD productivity and economic indicator with R2 of 0.67 and 0.72, respectively.