Research on Intelligent Production Optimization of Low-Permeability Tight Gas Wells

Yi Zhang, Xin Li, Shengguo Yang, Kewen Qiang, Bin Zhang, Jie Liu, Qiansheng Wei, Rui WangSymmetry 2025 · DOI

TL;DR

This paper introduces a novel hybrid deep learning model (P-C-L-A) that integrates Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM), and an attention mechanism, enhanced with physical constraints derived from gas well decline laws. This approach significantly improves the accuracy of production prediction for low-permeability tight gas wells, outperforming traditional methods and offering a more robust framework for intelligent production optimization throughout a well's lifecycle.

What Problem Does This Paper Address?

The efficient production of low-permeability tight gas wells is crucial for meeting energy demands, but it's hampered by complex geological conditions, reservoir heterogeneity, and the increasing number of low-yield wells. Traditional production prediction methods struggle with these complexities, leading to suboptimal management and reduced economic benefits. Specifically, accurately predicting the production capacity of these wells, especially in their early stages and throughout their lifecycle, remains a significant challenge. The paper aims to bridge this gap by developing an intelligent system that can accurately predict production and optimize operational strategies for these challenging reservoirs.

Methods

The core of this research is the P-C-L-A (Physical Constraint-Convolutional Neural Network-Long Short-Term Memory-Attention) model. This hybrid deep learning architecture combines:

  • CNNs: To capture local spatial features within the production data (e.g., pressure, water production).
  • LSTM: To model the temporal dependencies and sequential nature of production data over time.
  • Attention Mechanism: To allow the model to focus on the most critical time steps and features, enhancing its ability to handle complex dynamics and potential anomalies.

A key innovation is the integration of physical constraints derived from established gas well decline laws (Arps and SEPD models). These constraints are embedded into the loss function during training. This "dual-driven" approach ensures that the model's predictions not only align with historical data but also adhere to fundamental physical principles, improving interpretability and accuracy.

For production optimization, a two-stage strategy is proposed:

  1. Initial Static Similar Production Allocation: For new wells with limited data, a K-nearest neighbor (KNN) algorithm is used to identify similar wells based on geological and engineering parameters. This similarity-based approach guides the initial production allocation.
  2. Historical Data-Driven Rolling Production Allocation: For wells in stable production, the P-C-L-A model is used for short-term production prediction, with dynamic optimization achieved through rolling updates of production data.

Datasets

The study utilizes dynamic production data from four tight gas wells in the eastern area of the Sulige gas field, Ordos Basin. This data includes six dynamic characteristics: casing pressure, tubing pressure, production time, daily water volume, and daily gas production. The data was normalized to a [0, 1] range using deviation standardization to mitigate the influence of varying units and scales.

Key Insights and Ideas

  • Physical Constraints are Crucial: Embedding physical laws (like decline curves) into a data-driven model significantly improves prediction accuracy and stability, moving beyond a "black box" approach.
  • Hybrid Models Excel: Combining CNNs for spatial feature extraction and LSTMs for temporal sequence modeling, further enhanced by an attention mechanism, creates a powerful framework for complex time-series prediction.
  • Attention for Robustness: The attention mechanism is vital for filtering noise and focusing on critical data points, especially when dealing with potential anomalies or complex production behaviors.
  • Two-Stage Optimization is Effective: A staged approach to production optimization, starting with similarity-based allocation for new wells and transitioning to data-driven rolling optimization for established wells, addresses the challenge of varying data availability across a well's lifecycle.
  • Weighting Factor Balance: The optimal balance between data fitting and physical constraints (achieved through a weighting factor in the loss function) is critical for model performance.

Key Results

  • The P-C-L-A model achieved a 24.41% reduction in root mean square error (RMSE) compared to the traditional BP neural network.
  • Ablation studies demonstrated that each component (physical constraints, attention, LSTM, CNN) contributes significantly to the model's performance, with LSTM showing the greatest impact.
  • The inclusion of physical constraints reduced prediction errors, with MAE and RMSE for test sets being consistently lower than the model without constraints.
  • The staged production allocation strategy showed promising results:
  • The K-nearest neighbor algorithm for initial allocation yielded an average relative error of 7.21% for 10 wells.
  • For stable production wells, the P-C-L-A model achieved an average RMSE of 0.1618 × 10⁴ m³ for 30-day predictions.
  • The optimal weighting factor for balancing physical constraints and data fitting was found to be 0.6.

Limitations and Future Work

  • Data Scarcity in Early Stages: While the KNN approach addresses initial data scarcity, the inherent limitations of early-stage data can still lead to prediction deviations.
  • Generalizability: While the model was tested on Sulige gas field data, further validation on diverse tight gas fields with varying geological characteristics would enhance its universality.
  • Computational Cost: Deep learning models, especially hybrid ones, can be computationally intensive, requiring significant resources for training.

Future work could focus on exploring more advanced physics-informed neural networks, investigating transfer learning techniques to improve generalization across different fields, and developing more efficient training strategies to reduce computational burden. The authors also suggest reconducting sensitivity analyses of weighting factors for different gas fields to optimize model performance universally.

Summary updated

BibTeX
@article{zhang2025research,
  author = {Yi Zhang and Xin Li and Shengguo Yang and Kewen Qiang and Bin Zhang and Jie Liu and Qiansheng Wei and Rui Wang},
  title = {{Research on Intelligent Production Optimization of Low-Permeability Tight Gas Wells}},
  journal = {Symmetry},
  year = {2025},
  doi = {10.3390/sym17081311},
}
Figure