Optimization of Machine Learning Approaches for Shale Gas Production Forecast

Successful production from shale gas formations has changed the energy outlook of North America. During such process, a significant amount of data related to production and operations has been accumulated, which can be analyzed with machine learning (ML) techniques for production forecasts and performance optimization. In this study, an optimization workflow is proposed for production forecast in shale gas reservoirs, focusing on mitigating the overfitting problem while maximizing its accuracy. More specifically, a new objective function is developed where a penalty term is added to mitigate the model's overfitting tendency. The kernel density estimate (KDE) method guarantees that a similar distribution is shared between the split training and testing subsets. Four machine learning algorithms, including XGBoost, LightGBM, CatBoost, and Random Forest, are utilized and further optimized by the sparrow search algorithm with the newly developed objective function. A stacking approach is finally established based on the optimized algorithm and selected meta-models. Results show that the R2 of all four models has increased to around 0.9 after the optimization process, where XGBoost and LightGBM perform slightly better than the rest. In addition, R2 of the training dataset for XGBoost was decreased from 0.999 to 0.937, while the R2 for the testing dataset was increased from 0.762 to 0.905. This indicates that the overfitting has been mitigated after the optimization with the gap between the training and testing datasets reduced from 0.237 to 0.032. Moreover, stacking the optimized ML algorithms can not further enhance its accuracy. The best performer with SVM as the meta-model can only increase the R2 to 0.910, yet the model complexity increases significantly.