Decoupling and predicting natural gas deviation factor using machine learning methods

Area of Science:

• Petroleum Engineering
• Chemical Engineering
• Data Science

Background:

• Accurate natural gas deviation factor (Z-factor) prediction is vital for reserve estimation, reservoir recovery, and pipeline transport.
• Traditional machine learning models struggle with generalization across diverse gas compositions and conditions, limiting prediction accuracy.

Purpose of the Study:

• To develop an efficient machine learning framework for accurate natural gas Z-factor prediction.
• To enhance the robustness and generalization capabilities of Z-factor prediction models.

Main Methods:

• Proposed a hybrid approach utilizing signal decomposition algorithms (VMD, EFD, EEMD) to decouple Z-factor into components.
• Applied traditional machine learning algorithms (SVM, XGBoost, LightGBM, ANN, BiLSTM) to predict each decomposed component.
• Evaluated the impact of decomposition methods and component numbers on model performance, with SVM and VMD showing best results.

Main Results:

• The proposed framework significantly improved prediction accuracy across multiple traditional machine learning algorithms.
• Achieved an average correlation coefficient > 0.99 and mean absolute percentage error < 0.83% on 10 diverse datasets.
• Demonstrated superior performance compared to models without decomposition, especially under varying gas components and thermodynamic conditions.

Conclusions:

• The hybrid machine learning framework offers a unified and robust solution for precise natural gas Z-factor prediction.
• This advancement benefits the natural gas industry by improving resource estimation and reservoir management.
• The method effectively captures Z-factor variations, enabling reliable predictions across various operational scenarios.