PulseExploreJournal ClubDebatesTrendingResearchersJournals
Instagram
HomeExploreJournal ClubTrending
Synapse
⌘+K
Synapse
May 13, 2026Processes1 citationsOpen Access

Prediction of Three-Dimensional In Situ Stress in Deep Coal Rocks Considering Heterogeneity and Physical Information

View Full Paper
BLBing LiYKYunwei KangPHPengcheng Hao

Key Points

  • This research aims to improve the prediction of three-dimensional in situ stress in deep coal reservoirs by integrating physical information with data-driven methods.
  • Integrated a LightGBM prediction model with physical constraints based on Huang’s model.
  • Used the K-means algorithm to categorize the reservoir into three groups to mitigate heterogeneity.
  • Conducted comparative experiments with BP neural networks, random forests, and Transformers.
  • Achieved a MAPE of 2.78% and R2 of 0.89 on the test set with the proposed model.
  • Demonstrated improved prediction accuracy over traditional data-driven and mechanistic models.
  • Verified the model's applicability to blocks with similar geological conditions through transfer experiments.

Abstract

Deep coalbed methane reservoirs are characterized by complex geological conditions, strong heterogeneity, and significant variations in in situ stress, posing challenges for accurate three-dimensional in situ stress prediction. To address the issues of strong dependence on rock mechanical parameters in traditional physical models, as well as the lack of physical constraints and poor generalization capability under small-sample conditions in purely data-driven methods, this paper proposes a LightGBM prediction model that integrates physical information and data clustering. A total of 1289 fracturing clusters in the DJ block are selected as the research objects. First, the K-means algorithm is used to divide the reservoir into three categories to reduce the impact of heterogeneity. Then, a LightGBM model is constructed for each category, and physical constraints based on Huang’s model and stress–gravity equilibrium are incorporated into the loss function to ensure that the prediction results conform to mechanical laws. Taking the fracturing clusters in Category I as an example, the proposed model achieves an MAPE of 2.78% and an R2 of 0.89 on the test set. Comparative experiments show that the proposed model outperforms BP neural networks, random forests, and Transformers in prediction accuracy. Ablation experiments verify the independent contributions and synergistic effects of the clustering module and the physical information constraints. Transfer experiments demonstrate that the model has good applicability to blocks with similar geological conditions. This study provides an effective method for predicting in situ stress in deep coalbed methane reservoirs, balancing accuracy and physical interpretability.

Ask AI
Helpful
Bookmark
Share
View Full Paper

Cite This Study

Li et al. (2026) studied this question.

synapsesocial.com/papers/6a03cbe01c527af8f1ecfa5bhttps://doi.org/10.3390/pr14101535
Ask AI
Helpful
Bookmark
Share
View Full Paper