Low-rank coalbed methane production capacity prediction method based on time-series deep learning

Coalbed methane (CBM), a novel clean energy source, significantly contributes to reforming the energy supply structure, advancing carbon neutrality, and achieving peak carbon emissions. A precise production forecasting method is essential for understanding coalbed methane development and improving extraction efficiency. This paper proposes an improved multi-channel LSTM production forecasting model considering the influence of production dynamic factors, based on cutting-edge deep learning technology and widely applied production decline curve analysis methods. The model is experimentally tested on CBM wells in the Surat Basin of Australia. In the data processing segment of the model, we have improved the 3σ rule for handling outliers within a normal distribution. This enhancement maintains data integrity, achieves superior denoising, increases prediction accuracy, and significantly reduces model learning time. The test results demonstrate that the improved model exhibits excellent predictive performance for CBM wells, with 87.3 % of wells having a prediction accuracy exceeding 85 % overall. Additionally, the comprehensive multi-channel LSTM model identifies dynamic controlling factors in CBM production by analyzing variations in prediction accuracy with different dynamic factor combinations. This study provides new insights and scientifically rational references for production forecasting in CBM extraction. It's also crucial for effective production strategies and process optimization.