Comparison of different machine learning algorithms for predicting the SAGD production performance

Abstract

Steam Assisted Gravity Drainage (SAGD) is a typical thermal recovery process consisting of an upper horizontal injector and a lower horizontal producer, which heats and removes bitumen from oil sands deposits by steam. Serious research has been conducted on data-driven models to analyze the cumulative production performance of the SAGD process, and one of the most common machine learning methods utilized is the Artificial Neural Network (ANN). It is important to test other machine learning methods like Extreme Gradient Boosting (XGBoost) and Light Gradient Boosting Machine (LightGBM) in different conditions of data samples. In this paper, a series of SAGD models based on typical oil sands reservoir properties and operational conditions were constructed. Three different data groups were simulated by the Computer Modelling Group (CMG) thermal software STARS. Three different machine learning methods, including ANN, XGBoost, and LightGBM were constructed to calibrate a relationship between the input parameters and the output parameters in the different simulated data groups. A series of final models were tested and compared. The conclusion shows that data-driven training improves as the number and randomness of the data samples increase, and the LightGBM has the best prediction performance with such data samples. The ANN model may be the best choice with data samples of the worst randomness.