Unified Unit-Wise and Plantwide Monitoring: Application in Early Detection of Gas Flare Event

Gas flaring is a prevalent practice in various industrial processes, primarily instituted as a safety measure to relieve overpressure from storage vessels or pipelines, thereby enhancing process safety. However, while it bolsters operational safety, gas flaring is also a significant source of greenhouse gas emissions. Mitigating its environmental impact necessitates the early detection of flare events and prompt corrective actions. This paper introduces a novel integrated approach for the early detection of gas flare events, surpassing existing state-of-the-art methodologies. The proposed approach considers both unit-wise and plantwide dynamics to provide a more comprehensive approach to monitoring the process. The unit-wise monitoring is performed based on probabilistic slow feature analysis (PSFA), which is a linear dynamic latent variable model. This work implements PSFA in a moving window framework to make it adaptive to the time-varying process dynamics of each unit. Finally, the plantwide dynamic latent variables are extracted using a novel variational autoencoder-based architecture. The efficacy of the proposed algorithm is illustrated through a real-world case study from a refinery, showcasing its superior performance in terms of accuracy and reliability.