Optimized Flare Performance Analysis Through Multi-Modal Machine Learning and Temporal Standard Deviation Enhancements

Flaring is a routine practice in the upstream gas industry to dispose of waste gases, but its efficiency can drop significantly under non-ideal conditions such as crosswinds, over-aeration, or over-steaming. These inefficiencies lead to incomplete combustion, producing harmful substances like carbon monoxide and unburned methane, which contribute significantly to global warming. Current solutions for monitoring flare efficiency are often complex or expensive, limiting their widespread adoption. This work introduces a novel framework for estimating flare combustion efficiency (CE) using a multi-modal machine learning architecture enhanced by a Temporal Standard Deviation (TSD) preprocessing technique. Our approach combines synchronized visual data with minimal field measurements (FM) for accurate efficiency estimation. We first extract sequential frames from an RGB video stream of flares and process them to extract TSD images, which essentially highlight the variability and dynamic changes in the combustion process. Next, we extract image feature representations from TSD images using the state-of-the-art Vision Transformer (ViT) and fuse them with experimentally selected FM data to create a comprehensive combustion dataset. A CatBoost regression model is then trained on this dataset to estimate the final CE. Our proposed framework is validated using real-world data from industrial flare upstream operations. The results demonstrate significant improvements in estimation accuracy and reliability compared to traditional methods, achieving an R-squared score of 94.77% with minimal FM data. This approach not only enhances the understanding of combustion dynamics but also offers a scalable, cost-effective solution for continuous flare monitoring and optimization.