Machine Learning-Based Research for Predicting Shale Gas Well Production

Nijun Qi, Xizhe Li, Zhenkan Wu, Yujin Wan, Nan Wang, Guifu Duan, Longyi Wang, Jing Xiang, Yaqi Zhao, Hongming ZhanSymmetry 2024 · DOI

TL;DR

This paper tackles the challenge of accurately predicting shale gas well production (Estimated Ultimate Recovery - EUR) by identifying the most influential geological, engineering, and production factors. Using a combination of feature selection techniques and machine learning models, they developed an optimized XGBoost model that significantly improves prediction accuracy, offering a more efficient approach for reservoir development.

What Problem Does This Paper Address?

Accurately predicting the Estimated Ultimate Recovery (EUR) of shale gas wells is crucial for effective reservoir development and economic assessment. However, this is a complex task due to the intricate nature of shale reservoirs, characterized by nanoscale pore structures, multi-phase flow, and fluctuating production. Traditional methods for EUR evaluation, such as empirical, numerical simulation, and analytical approaches, often face limitations in applicability, accuracy, and the handling of reservoir heterogeneity. The authors aim to overcome these limitations by leveraging machine learning to build a more robust and accurate prediction model, focusing on identifying and utilizing the most impactful influencing factors.

Methods

The study employs a multi-step methodology:

  • Feature Screening:
    • Pearson Correlation and Mutual Information: These statistical methods were used to initially assess the relationships between 31 potential influencing factors and EUR, helping to identify highly correlated and redundant features.
  • Random Forest - Recursive Feature Elimination (RF-RFE): This algorithm was employed to systematically rank and select the most important features from the remaining set, aiming to reduce dimensionality and improve model performance. RF-RFE is highlighted for its ability to provide a quantitative measure of feature importance and its suitability for datasets with many features.
  • Prediction Models: Four machine learning models were trained and evaluated:
  • Single Models: Multiple Linear Regression (MLR) and Support Vector Machine (SVM).
  • Integrated Models: Random Forest (RF) and Extreme Gradient Boosting (XGBoost). XGBoost is noted for its systematic and efficient implementation of gradient boosting, incorporating regularization and automatic handling of missing values.
  • Model Optimization and Evaluation:
    • Bayesian Optimization: Used to optimize hyperparameters for the integrated models (RF and XGBoost) and single models (SVM), aiming to find the best model configuration.
  • Performance Metrics: Mean Absolute Error (MAE), Mean Squared Error (MSE), and Coefficient of Determination (R²) were used to assess prediction accuracy. Taylor diagrams and scatter plots were also utilized for visual comparison of predicted and observed values.

Datasets

The study utilized a dataset comprising 31 features from 200 shale gas wells in the Weiyuan block of the Sichuan Basin. The data was pre-processed using normalization to the [0, 1] interval to improve model performance. The dataset was split into an 80% training set and a 20% test set. No new benchmark datasets were introduced; the focus was on analyzing existing well production data.

Key Insights and Ideas

  • Dominating Factors are Key: The research strongly emphasizes that not all features are created equal. Identifying and using only the most influential factors (the "dominating factors") significantly improves prediction accuracy compared to using all available data. This highlights the importance of feature engineering and selection in machine learning applications for complex geological systems.
  • Integration Models Shine: Ensemble methods like XGBoost and Random Forest outperformed single models (MLR and SVM). This suggests that combining multiple base learners can effectively reduce bias and variance, leading to more robust predictions.
  • XGBoost's Superiority: The XGBoost model emerged as the top performer, demonstrating the highest R² and lowest MAE/MSE. Its ability to handle complex interactions and its inherent regularization likely contribute to its success in this domain.
  • Geology, Engineering, and Production Interplay: The identified dominating factors span all three categories, underscoring that a holistic understanding of well performance requires considering geological potential, the effectiveness of well completion (drilling and fracturing), and the subsequent production dynamics.

Key Results

  • Six Dominating Factors Identified: The RF-RFE algorithm identified six key factors influencing shale gas well EUR:
  • Geology: High-quality reservoir thickness.
  • Engineering: High-quality reservoir length drilled, fracturing fluid volume, proppant volume, and fluid volume per length.
  • Production: 360-day flowback rate.
  • XGBoost Model Performance: The optimized XGBoost model achieved an R² of 0.804 on the test set, with an MAE of 0.130 and MSE of 0.024.
  • Improved Accuracy with Feature Selection: Using the six dominating factors as input improved the R² values for all models compared to using the original 31 features. For XGBoost, the R² increased by 3.6%.
  • Correlation: The XGBoost model showed a high correlation coefficient of 0.9 between predicted and observed values.

Limitations and Future Work

  • Sample Size: The authors acknowledge that the sample size of 200 wells, while sufficient for demonstrating the methodology, is limited. They suggest that a larger dataset would likely lead to even higher prediction accuracy, as machine learning models generally benefit from more data.
  • Data Confidentiality: The raw data is not publicly available due to confidentiality restrictions, which limits external validation and replication by other researchers.
  • Future Work: The paper suggests developing a larger sample database for future studies. The optimized XGBoost model, with its high accuracy and reduced data requirements, is presented as a novel and efficient tool for predicting shale gas well production, paving the way for further research in this area.

Summary updated

BibTeX
@article{qi2024machine,
  author = {Nijun Qi and Xizhe Li and Zhenkan Wu and Yujin Wan and Nan Wang and Guifu Duan and Longyi Wang and Jing Xiang and Yaqi Zhao and Hongming Zhan},
  title = {{Machine Learning-Based Research for Predicting Shale Gas Well Production}},
  journal = {Symmetry},
  year = {2024},
  doi = {10.3390/sym16050600},
}
Figure