Small-Sample Production Prediction of Fractured Wells Using Multitask Learning

Domestic and foreign scholars have conducted extensive research on applying machine learning to post-fracture production prediction in recent decades and made great achievements in Bakken, Eagle Ford, Marcellus, and other large-scale oil and gas fields. However, few studies focus on small-sample production prediction of fractured wells, which is urgently needed in small-scale and newly developed reservoirs. In this work, we propose a novel small-sample production prediction framework based on multitask learning (MTL), including multitype data collection, task selection, data preprocessing, model training, and multitask production prediction. As for the trained model, feature extraction is first used through the deep hybrid network (DHN) by fully leveraging available multitype data such as numerical, sequence, and image. Then a multitask module based on the cross-stitch network (CSN) is integrated to automatically determine the information sharing degree of multiple production prediction tasks. In this way, data augmentation and model regularization arc indirectly realized to handle the overfitting problem caused by insufficient training data. Finally, the proposed framework is applied to a small-sample field case with 59 fractured horizontal wells in northwest China. The comparison results show that the developed MTL-based model performs better than traditional single-task models in both prediction accuracy and learning efficiency. It provides an accurate and efficient tool for small-sample production prediction and can be used for reference to other small-sample domains in the petroleum industry.