Machine learning techniques to increase the performance of indirect methane quantification from a single, stationary...

Area of Science:

• Environmental Science
• Data Science
• Chemical Engineering

Background:

• Accurate quantification of methane emissions from natural gas infrastructure is crucial for environmental monitoring and regulatory compliance.
• Current indirect quantification techniques (IQTs) like Other Test Method (OTM) 33A and Eddy Covariance (EC) offer advantages in measurement frequency and resource efficiency but suffer from accuracy and repeatability issues.
• Direct measurement methods are accurate but resource-intensive, limiting their widespread application.

Purpose of the Study:

• To develop and evaluate a novel approach for improving the accuracy of single-sensor IQTs for methane emission quantification.
• To integrate machine learning (ML) models, specifically random forest (RF) and neural network (NN), with OTM and EC data.
• To enhance ML model performance through feature reduction and hyper-parameter tuning for more reliable methane emission estimates.

Main Methods:

• Combined data from OTM 33A and EC measurements.
• Developed and trained RF and NN ML models using the combined IQT data.
• Applied feature reduction and hyper-parameter tuning to optimize ML model performance.
• Compared ML model results against traditional quantification methods and direct measurements.

Main Results:

• The NN and RF models demonstrated significant improvements over the default OTM, with average accuracy enhancements of 44% and 78%, respectively.
• When compared to OTM estimates with low Data Quality Indicators (DQIs), the RF and NN models reduced standard deviation errors from ±66% to ±13% and ±34%, respectively.
• The models achieved high precision, with 93% (RF) and 85% (NN) of estimates falling within ±50% of the known methane release rate.

Conclusions:

• The proposed ML-enhanced single-sensor IQT approach substantially improves the accuracy and reliability of methane emission quantification from natural gas infrastructure.
• This method can be deployed at well sites to increase confidence in reported emissions, reducing false positives for site evaluations.
• Further development with expanded datasets and diverse deployment scenarios can enhance the ability to monitor temporal emissions and bridge the gap between bottom-up and top-down studies.