What question did this study set out to answer?

This research aims to enhance coalbed methane production strategies through AI-based optimization techniques.

April 8, 2026

Optimization of Multiwell Coalbed Methane Production Based on Surrogate Model and Proximal Policy Optimization Algorithm

Key Points

This research aims to enhance coalbed methane production strategies through AI-based optimization techniques.
Developed a high-fidelity surrogate model using a dual-stream neural network architecture.
Utilized long short-term memory (LSTM) and graph convolutional networks (GCN) for feature extraction and spatial modeling.
Integrated the surrogate model with a reinforcement learning agent employing proximal policy optimization (PPO).
Achieved a predictive R2 of 0.9897 with minimal errors on unseen data.
Increased cumulative gas production by 13.19% while limiting water drainage costs to 5.36%.
Resulted in a 12.13% increase in the project's net present value (NPV).

Abstract

Summary Optimizing production strategies for multiwell coalbed methane (CBM) fields presents a significant challenge due to complex spatiotemporal reservoir dynamics and the high-dimensional, interdependent nature of well control decisions. To address this challenge, we propose a comprehensive, two-stage artificial intelligence (AI) framework. First, we develop a high-fidelity surrogate model based on a parallel dual-stream spatiotemporal neural network. This architecture uniquely decouples feature extraction, using a long short-term memory (LSTM) branch to capture individual well dynamics and a graph convolutional network (GCN) branch to model instantaneous spatial correlations. The model demonstrates exceptional accuracy in predicting production, achieving a predictive R2 of 0.9897, a root mean square error of 126.49 m3/month, and a mean absolute error of 43.52 m3/month on unseen data. Second, this surrogate model is integrated as the environment for a sophisticated reinforcement learning (RL) agent tasked with optimizing the multiwell production strategy. The agent is powered by proximal policy optimization (PPO), chosen for its stability in continuous control landscapes, and features two key innovations—a self-attention mechanism in its policy network to dynamically infer interwell dependencies and generalized state-dependent exploration (gSDE) for efficient policy discovery. Through comprehensive ablation studies, we demonstrate the superiority of the combined architecture. The final trained agent discovers a production strategy that increases cumulative gas production by 13.19% while only increasing water drainage costs by 5.36% compared with the historical baseline. This translates to a 12.13% increase in the project’s net present value (NPV). This work establishes an end-to-end AI framework that significantly enhances production efficiency and provides a robust paradigm for dynamic reservoir management.

اسأل الذكاء الاصطناعي

Bookmark