Adaptive anomaly detection-based liquid loading prediction in shale gas wells

Abstract

Liquid loading is one of the main factors that can severely impede shale gas production. It is crucial for efficient operation of gas production to forecast the liquid loading in real time. Most of the existing techniques for liquid loading prediction are based on physical modeling, which are generally limited by human knowledge of the underlying mechanism of liquid loading formation. In this study, we propose a data-driven method for liquid loading prediction in shale gas wells based on deep anomaly detection technology. In particular, we employ deep neural networks to learn the normal behavior of a well and test a given sequence for anomaly by measuring the sequence reconstruction errors of the neural network with adaptive threshold. Unlike existing physical model-based methods, our method makes no assumptions about the cause of liquid loading, while the prediction results are merely determined by historical records. To verify the effectiveness of our method, we conduct experiments on real shale gas production data coming from 73 shale gas wells and the precision of method is 85.71%. The experimental results demonstrate that our method surpasses existing physical modeling based methods at prediction accuracy and timeliness.