Well-Production Forecasting Using Machine Learning with Feature Selection and Automatic Hyperparameter Optimization

Well-production forecasting plays a crucial role in oil and gas development. Traditional methods, such as numerical simulations, require substantial computational effort, while empirical models tend to exhibit poor accuracy. To address these issues, machine learning, a widely adopted artificial intelligence approach, is employed to develop production forecasting models in order to enhance the accuracy of oil and gas well-production predictions. This research focuses on the geological, engineering, and production data of 435 fracturing wells in the North China Oilfield. First, outliers were detected, and missing values were handled using the mean imputation and nearest neighbor methods. Subsequently, Pearson correlation coefficients were utilized to eliminate linearly irrelevant features and optimize the dataset. By calculating the gray correlation degrees, maximum mutual information, feature importance, and Shapley additive explanation (SHAP) values, an in-depth analysis of various dominant factors was conducted. To further assess the importance of these factors, the entropy weight method was employed. Ultimately, 19 features that were highly correlated with the target variable were successfully screened as inputs for subsequent models. Based on the AutoGluon framework, model training was conducted using 5-fold cross-validation combined with bagging and stacking techniques. The training results show that the model achieved an R2 of 0.79 on the training set, indicating good fitting ability. This study offers a promising approach for the development of oil and gas production forecasting models.