An enhanced prediction framework for coalbed methane production incorporating deep learning and transfer learning

Abstract

Coalbed methane (CBM) is one of the most promising clean energy sources, and its production forecast is of vital strategic significance for the sustainable development. Traditional numerical simulation encounters the curse of dimensionality when applied to multi-scale reservoir modeling. Conventional machine learning faces challenges in modeling under small-sample condition. This study leverages a bidirectional long short-term memory (Bi-LSTM) to better capture the complex dynamic data characteristics of CBM. Moreover, we incorporate transfer learning to generalize the implicit laws learned from 1205 vertical wells to the modeling of horizontal wells, establishing an enhanced prediction framework that can be adapted to different well types. The findings confirm that Bi-LSTM has more powerful memory capability for CBM dynamic data with prediction errors below 200 m³, which can improve the prediction accuracy by about 15% compared to the traditional methods. The fusion of transfer learning can significantly enhance the prediction performance by approximately 10% for both low-producing, medium-producing, and high-producing horizontal wells, illustrating that the dynamic data of vertical wells located in the same block contain implicit features that are crucial guides to the modeling of horizontal wells. Notably, the enhanced architecture exhibits greater flexibility and adaptability under small-sample condition for complex production patterns.