Field dataset collected from Atuabo gas processing plant in Ghana for a specified period

Accurate quantification of isopentane flow is critical for operational accounting and environmental compliance at gas processing facilities, yet the absence of physical meters and unreliable mass-balance estimates often create significant data gaps. This study addresses this challenge at the Atuabo Gas Processing Plant in Ghana by developing a hybrid computational framework to estimate unmetered isopentane flow and its associated emissions. A validated steady-state Aspen HYSYS digital twin, benchmarked against 2,404 hourly field records, generated high-fidelity synthetic training data. An Artificial Neural Network (ANN) was subsequently trained on six key inlet process variables as a real-time surrogate model. Computational Fluid Dynamics (CFD) modelling in ANSYS Fluent established a site-specific flare combustion efficiency of 97.06%, which was integrated with the ANN output to quantify CO₂ emissions under an IPCC Tier 3 framework. The ANN achieved an R² of 0.794 and an RMSE of 168.16 kg/h, outperforming Support Vector Regression, Random Forest, and Gradient Boosting baselines. KernelSHAP analysis identified inlet gas flow rate and the feed–stripper temperature difference as the dominant predictors. Deployed on plant data, the system estimated a mean isopentane flow of 3,146 kg/h and a mean CO₂ emission rate of 9,193 kg/h (~80,533 tonnes/year). This work demonstrates that an ANN-based soft sensor anchored by physical simulation offers a scalable, cost-effective alternative to physical metering for real-time carbon accounting in gas processing facilities.