Author ORCID Identifier

https://orcid.org/0009-0002-8466-691X

Semester

Fall

Date of Graduation

2025

Document Type

Thesis

Degree Type

MS

College

Statler College of Engineering and Mineral Resources

Department

Mechanical and Aerospace Engineering

Committee Chair

ANDREW C. NIX

Committee Member

DEREK JOHNSON

Committee Member

VYACHESLAV AKKERMAN

Abstract

Abstract

ASSESSING AND PREDICTING THE IMPACT OF A RETROFITTED METHANE MITIGATOR ON NON-ROAD LARGE NATURAL GAS ENGINES USING ARTIFICIAL NEURAL NETWORKS

Augustine Thomas Ojo

Methane emissions from the oil and gas sector present a pressing environmental and economic challenge, given methane’s high global warming potential and the considerable volumes of unutilized gas lost from industrial processes. This study investigates the impact of a retrofitted methane mitigation system, the Methane Mitigator (M²), on fuel consumption and engine-out emissions in non-road, four-stroke natural gas-powered, lean-burn engines and develops predictive Artificial Neural Network (ANN) models to estimate the resulting engine-out emissions and fuel consumption. Key methane sources targeted for mitigation include reciprocating compressor vents (RC), pneumatic controllers (PC), and engine crankcase vents (CC), each repurposed as supplementary fuel streams.

Two training algorithms, Levenberg–Marquardt (LM) and Bayesian Regularization (BR), were employed to model six output parameters: carbon dioxide (CO2), methane (CH4), carbon monoxide (CO), nitrogen oxides (NOx), total hydrocarbons (THC), and fuel consumption. Model training utilized emissions data generated in West Virginia University’s emissions-testing laboratory, augmented with Caterpillar’s Gas Engine Rating Pro (GERP) database, enabling extrapolation to 16 untrained Caterpillar engine configurations (including G3520J, G3516J, G3512J, and G3508J models). These engines span displacements from 34.5L to 86.3L, covering medium- and large-size categories, and were analyzed under three load levels (50%, 75%, 100%). Under baseline conditions, the models achieved strong predictive accuracy across all emission components, with over 80% of predictions falling within ± 10% error margins across the tested engine configurations, sizes, loads, and speeds.

For the mitigated scenarios, model training was conducted exclusively on data from a single laboratory engine, presenting a common extrapolation challenge when applied to field engines of varying configurations. To address this, the model architecture employed a feedforward artificial neural network (ANN) with ReLU activation (poslin) in the hidden layers and a linear (purelin) output layer. This configuration, combined with tailored training parameters, demonstrated high extrapolative capacity without further retraining.

When applied to untrained field engine configurations, the model achieved strong predictive performance across both the reciprocating compressor and engine crankcase mitigation scenarios. For the engine crankcase scenario, five of six output components were well-predicted, with CO₂ achieving less than 20% deviation from the baseline in 88.9% of test datapoints across medium (200–1000 BHP) and large (BHP>1000) engines. CH₄ and THC achieved 63.7% and 68.9% within this deviation band, respectively. Similarly, for the reciprocating compressor case, CO₂, CH₄, and THC predictions fell within the 20% deviation threshold for 91.1%, 40.0%, and 71.9% of the datapoints, respectively.

The approach provides a scalable and cost-efficient framework for estimating the emissions and fuel performance of retrofitted methane mitigation systems, without requiring extensive instrumentation or re-calibration for each engine deployment.

Download

Share

COinS