Volume 41: Energy Transitions toward Carbon Neutrality: Part IV

Comparative Study of Machine Learning Algorithms for Predicting Steam-assisted Gravity Drainage (SAGD) Production Performance Bin Wang, Shijun Huang, Fenglan Zhao, Yue Chen, Xinhan Fan



Steam-assisted gravity drainage (SAGD) is one effective and well-established technology for recovering heavy oil and bitumen resources. Extensive research has been conducted on data-driven models to evaluate the production performance of the SAGD process. The artificial neural network (ANN) is a commonly used machine learning method. However, it is crucial to explore other machine learning methods such as Symbolic Regression (SR), Extreme Gradient Boosting (XGBoost) and Random Forest (RF) using field data. In this study, firstly, a data set consisting of thirteen input/output attributes describing production-related properties and production characteristics was extracted from Long Lake field data. Secondly, three different machine learning methods, including Neural Networks (ANN), Random Forest (RF), Extreme Gradient Boosting (XGBoost) and Symbolic Regression (SR), were employed to establish a relationship between the input and output parameters in the different data sets. Subsequently, a range of models were created, evaluated, and compared. Furthermore, the impact of two feature scaling methods, namely standardization and normalization, on the accuracy of a series of prediction models was explored. Lastly, the sensitivity of the input parameters was analyzed. Analysis of the forecasting results obtained from different models leads to the following conclusions. The study found that standardization and normalization significantly enhance the performance of the artificial neural network model, with standardization being more effective. However, the impact of data scaling on integrated learning models (random forest and extreme gradient boosting tree) is minimal. Interestingly, for models based on symbolic regression algorithms, not using data scaling yields the best results. Both artificial neural network and symbolic regression algorithms demonstrate significant advantages and are suitable for predicting SAGD production. However, the symbolic regression model can derive analytical expressions that describe the input-output relationship, which are easy to interpret and apply. This suggests that symbolic regression algorithms may be preferable to artificial neural network algorithms. The top five factors with the greatest impact on cumulative oil production at the end of the main production stage in SAGD, ranked in descending order of importance, are CSI2 (cumulative steam injection volume at the end of the main production stage), LE (effective length of the horizontal well section), MSIP (mean steam injection pressure during the main production stage), HPVH (height of the hydrocarbon pore volume), and H (effective thickness). The methods proposed in this article are helpful for establishing an intelligent energy management system for SAGD (Steam-Assisted Gravity Drainage) development in heavy oil fields. The system can support decision-making processes by providing accurate forecasting and predictive analytics.

Keywords Steam-Assisted Gravity Drainage (SAGD), Machine learning, Artificial Neural Networks, Symbolic Regression, Random Forest, Extreme Gradient Boosting

Copyright ©
Energy Proceedings