An EA-LSTM–COMSOL Coupled Inversion Framework for Pressure-Field Reconstruction in Coalbed Methane Reservoirs

Abstract

Sparse underground pressure monitoring limits the direct characterization of full-field pressure distributions during coalbed methane (CBM) extraction, while conventional numerical simulations often require repeated parameter adjustments. This study proposes an intelligent inversion framework that couples an enhanced attention-based long short-term memory (EA-LSTM) model with a COMSOL-based forward model. Coal Seam No. 3 at Xinjing Mine was selected as the engineering case study. The vertical stress correction coefficient, reservoir pressure, extraction pressure, and initial permeability were treated as inversion parameters, while the working-face advance distance was introduced to represent the mining stage. A multiparameter sample library was constructed using an orthogonal design and segmented advance conditions, and pressure responses from five extraction boreholes were used as inversion constraints. The proposed model achieved a mean squared error of 0.013 and a coefficient of determination of 0.987, outperforming nine benchmark machine-learning and deep-learning models. The relative errors between the inverted and measured borehole pressures ranged from 0.41% to 5.07%, with a value of 1.84%. The coupled framework improves the efficiency of boundary-parameter identification and enables rapid reconstruction of pressure fields in a coalbed methane reservoir.