Analysis of dynamic production characteristics and identification of segmentation nodes in decline models for deep coalbed methane using physics-informed neural networks

Abstract

Deep coalbed methane (CBM) offers clean, sustainable energy, but its low-carbon potential and resource value require enhanced production. Conventional data-driven models lack physical interpretability and often miss key reservoir dynamics. This study presents a physics-informed neural network (PINN) long short-term memory (LSTM) model that embeds decline-curve analysis (DCA) and the Langmuir adsorption–desorption equation as physical constraints. Production time-series from six Ordos Basin wells were used for training and validation. Hyperbolic, stretched-exponential production decline (SEPD), Duong, and Langmuir constraints were systematically evaluated and linked to production–extraction mechanisms. A time-varying coupling strategy divides the production cycle into segments, applying the optimal decline model in each stage. Compared to a standard LSTM, the PINN-LSTM delivers higher accuracy, robustness, and interpretability. Early-stage fluctuating wells gain from Langmuir regularization, while mid-to-late-life wells see improved precision under hyperbolic and Duong constraints. The coupling strategy reduces prediction error by 10 % and raises R² by 0.11. Temporal segmentation pinpoints transition nodes and captures operational fluctuations. These results guide intelligent CBM production control and future model enhancements.