A data-driven workflow for prediction of fracturing parameters with machine learning

Abstract

In the realm of unconventional reservoir hydraulic fracturing design, the conventional optimization of mechanistic model parameters is a time-consuming process that impedes its responsiveness to the swift demands of on-site development. This study, rooted in Xinjiang oilfield data, delves into the utilization of machine learning methods for extensive field data. The research systematically elucidates the training and optimization procedures of a production forecasting model, achieving effective optimization of hydraulic fracturing design parameters. By employing polynomial feature cross-construction generate composite features, feature filtering is performed using the maximal information coefficient. Subsequently, wrapper-style feature selection techniques, including ridge regression and decision trees, are applied to ascertain the optimal combinations of model input parameters. The integration of stacking during model training enhances performance, while stratified K-fold cross-validation is implemented to mitigate the risk of overfitting. The ultimate optimization of hydraulic fracturing design parameters is realized through a competitive learning particle swarm algorithm. Results indicate that the accuracy of the data-driven production forecasting model can reach 85%. This model proficiently learns patterns from mature blocks and effectively applies them to optimize new blocks. Furthermore, expert validation confirms that the optimization results align closely with actual field conditions.