In a paper published in the journal Scientific Reports, researchers addressed the challenge of estimating gas viscosity, which is crucial for petroleum engineers. They employed seven machine learning (ML) techniques, including artificial neural network (ANN) and random forest (RF), on over 4304 datasets to predict methane (CH4), nitrogen (N2), and natural gas mixture viscosities.
The ANN, RF, and gradient boosting (GB) models achieved high precision with an R² of 0.99, while Linear Regression and Nu support vector regression (NuSVR) performed poorly. These ML models provide a cost-effective, fast alternative to experimental measurements, benefiting research and industry in approximating gas viscosities under various conditions.
Related Work
Past work has primarily focused on gas viscosity modeling using ANNs and adaptive neuro-fuzzy inference systems (ANFIS). Limited studies have applied ML techniques specifically for methane and nitrogen viscosity prediction. Managing the complexity of integrating new digital technologies with existing infrastructure and processes is a primary challenge in industrial digital transformation (DTI).
Another significant hurdle involves addressing heightened cybersecurity risks accompanying the increased digital connectivity required for DTI initiatives. Overcoming these challenges requires strategic planning, robust cybersecurity measures, and effective integration strategies to ensure seamless adoption of digital technologies in the industry.
Statistical Analysis of PVT Experiments
This article employs statistical methods to analyze published PVT experiments on CH4, N2, and gas mixtures under standard and severe conditions of high temperature and pressure. The study aims to elucidate the influence of critical factors such as pressure, temperature, and gas density on viscosity during these experiments.
Initial data collection involved compiling and comprehensively explaining experimental pressure-volume-temperature (PVT) data and results through histograms integrating pressure, temperature, gas density, and viscosity for CH4, N2, and gas mixtures. Additionally, the study established relationships between viscosity and the influencing variables through sensitivity analysis, analysis of variance (ANOVA), factorial design, contour plots, and main effects plots.
The novelty of this research lies in its application of advanced ML models to predict viscosities across a broad spectrum of conditions using over 4304 real experimental datasets. Additionally, the study introduces three new ANN-based correlations specifically tailored for CH4, N2, and gas mixture viscosities. The research offers a comprehensive evaluation and prediction framework unmatched in previous literature by employing seven distinct ML algorithms, including Linear Regression, DT, RF, GB, NuSVR, ANN, and K-NN.
In addition to modeling and predicting viscosity, this research enhances natural gas operations, processing efficiency, and reliability. The study improves the understanding of viscosity dynamics under varying conditions by leveraging advanced statistical analyses and ML techniques. It offers practical insights for optimizing gas transport, storage, and utilization strategies.
The findings underscore the importance of accurate viscosity predictions in ensuring operational safety, cost-effectiveness, and environmental sustainability across the natural gas industry. This comprehensive approach benefits researchers and engineers and provides valuable tools for decision-making and innovation in energy sector applications.
Advanced Viscosity Prediction
This study employed ANNs to model the viscosity of methane, nitrogen, and gas mixtures. They excel in learning and memory, leveraging training data, and were fine-tuned through diverse setups, including different transfer functions and varying neuron counts in hidden layers.
For methane, an ANN with a logistic sigmoid transfer function and eight neurons in the hidden layer achieved high precision (R2 = 0.99918), demonstrating strong predictive capability. Similarly, models for nitrogen and gas mixtures utilized deeper architectures with multiple hidden layers and achieved high accuracy (R2 = 0.997 and 0.9998, respectively), showcasing robust performance across varying pressure and temperature conditions.
Moreover, the models demonstrated exceptional alignment with experimental data, which is evident from regression plots and minimal error distributions observed across training, validation, and testing datasets. The study highlighted the potential of ANNs to efficiently predict viscosity under diverse environmental conditions, offering a faster alternative to traditional experimental methods.
While promising, these models are most effective within the bounds of their training data, necessitating caution when extrapolating to conditions not covered during training to maintain predictive accuracy. Ongoing advancements in data collection and model refinement could further enhance the reliability and applicability of these ANN-based approaches in predicting viscosity for industrial and scientific applications.
Conclusion
To summarize, this study utilized ML models to predict methane, nitrogen, and gas mixture viscosities under high-temperature and high-pressure conditions. A comprehensive database of 4304 experimental data sets was compiled from literature sources.
Novel correlations were developed using ANN, achieving high accuracy with R2 values of 0.99 for testing data sets. ANN, GB, and RF demonstrated superior performance in viscosity prediction compared to other models like KNN and Linear Regression. This research underscored the efficacy of data-driven ML approaches in enhancing the modeling of natural gas operations.
Journal reference:
- Gomaa, S., et al. (2024). Machine learning prediction of methane, nitrogen, and natural gas mixture viscosities under normal and harsh conditions. Scientific Reports, 14:1, 15155. DOI:10.1038/s41598-024-64752-8, https://www.nature.com/articles/s41598-024-64752-8