Use of artificial intelligence for emission monitoring and control in hydrocarbon production

Abstract

The article addresses the problem of greenhouse emissions, primarily methane and CO₂, during hydrocarbon production and aims to develop and test an integrated artificial-intelligence-based platform for continuous monitoring and proactive management of emissions, capable of replacing reactive and discrete approaches with continuous intelligent control. The empirical base covers 75 sites (50 onshore and 25 offshore) from 2018 to 2022, combining satellite data from Sentinel 5P and Landsat 8/9, UAV aerial surveys with LIDAR/hyperspectral equipment, telemetry from more than 15,000 IoT sensors (CH₄, CO₂, VOC, pressure, temperature, vibration, acoustics), and SCADA/MRO datasets totaling over 5 PB; preprocessing includes stream synchronization, cleaning, and normalization. The methodological framework combines CNNs (U-Net, Faster R-CNN) for segmentation/detection of gas plumes, LSTM for time series analysis, as well as XGBoost/LightGBM and Random Forest for anomaly and failure prediction with cross-validation and testing at test sites simulating leaks. The following results were obtained: the accuracy of unorganized emissions detection reached 98.72% with 1.15% false alarms and an average detection time of 0.25 h, which reduces the reaction window from 24 hours to 15 minutes and increases the efficiency of maintenance teams; after the platform’s implementation, CH₄ emissions decreased by 45.64% (versus 8.31% under traditional practices), CO₂ by 21.89%, VOC by 35.17%, and the average annual reduction exceeded 12 thousand tons of CO₂ equivalent per facility. The economic analysis notes a reduction in product loss costs from $4.88 million to $1.35 million per year and in environmental fees from $2.15 million to $0.44 million per year, alongside an increase in planned preventive maintenance costs to $2.75 million per year, resulting in a total net effect of $5.09 million per facility (payback less than 2 years). Predictive modules provide a 30-day forecasting horizon for compressors (Precision 94.11%; Recall 91.56%), 14 days for valve systems (97.43%; 92.81%), and 21 days for pumping stations (95.89%; 94.22%). The discussion highlights the synergy of multimodal fusion analytics, integration with PCS/EAM systems, and prospects for applying transformers, federated learning, digital twins, and robotic tools for rapid leak mitigation.