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

Researchers are searching for ways to better quantify methane emissions from natural gas infrastructure. Current indirect quantification techniques (IQTs) allow for more frequent or continuous measurements with fewer personnel resources than direct methods but lack accuracy and repeatability. Two IQTs are Other Test Method (OTM) 33A and Eddy Covariance (EC). We examined a novel approach to improve the accuracy of single sensor IQT whereby the results from both OTM and EC were combined with two machine learning (ML) models, a random forest (RF) and a neural network (NN). Then, models were enhanced with feature reduction and hyper-parameter tuning and compared to traditional quantification methods. The NN and RF improved upon the default OTM by an average of 44% and 78%, respectively. When compared to traditional OTM estimates with low Data Quality Indicators (DQIs), RF and NN models reduced 1σ errors from ±66% to ±13% and ±34%, respectively. Models also reduced the standard deviation of estimates with 93% and 85% of estimates falling within ±50% of the known release rate. This approach can be deployed with single sensor systems at well sites to improve confidence in reported emissions, reducing the number of anomalous overestimates that would trigger unnecessary site evaluations. Additional improvements could be realized by expanding training datasets with more methane release rates. Further, deployment of such models in a variety of situations could enhance their ability help close the gap between bottom-up inventory and top-down studies by enabling continuous monitoring of temporal emissions that could identify with improved confidence, atypically higher emissions. Accurate remote single sensor systems are key in developing an improved understanding of methane emissions to enable industry to identify and reduce methane emissions.