2024 AIChE Annual Meeting
(372s) Data-Driven Model for Boil-Off Gas Estimation in a Liquefied Natural Gas Regasification Terminal
Authors
To build a data-driven model, the operations of a process plant are first categorized into two main modes: steady states and transitions, as the plant’s control operations will differ depending on its current state. A steady state represents a period of continuous operation under fixed operating conditions, where the process operates at a steady state, and its variables fluctuate within a narrow range. In contrast, transitions involve significant changes or discontinuities in plant operations, such as altering setpoints, activating or deactivating equipment, or adjusting manual valves. These transitions introduce additional complexity to monitoring tasks due to the significant changes in observable plant variables compared to steady-state operations. This distinction between steady-state and transitions underscores the challenges associated with monitoring and control in dynamic process environments (Srinivasan et al., 2004). In the LNG regasification terminal, where no transfer of LNG occurs between the LNG carrier and the storage tank, but a steady send-out of LNG from the tank continues, is termed the holding mode. The identification of this holding mode apart from the transition is necessary within the plant to ensure accurate prediction of BOG.
Several methods have been previously suggested for identifying steady-state and transitions in agile chemical plants. Statistical techniques such as the t-test, F-test, and wavelet-based approach have been utilized in literature for steady-state identification. Furthermore, Srinivasan et al. (2005) developed a trend-based technique for segmenting modes and transitions using critical variables. However, these approaches are univariate, making their implementation in large-scale processes challenging. Hence, Seng et al. (2006) addressed the need for improved state identification by proposing a methodology for multivariate processes based on the approach proposed by Srinivasan et al. (2004). The criterion for transition identification is predicated upon a decision variable formed through the combination of multivariate scores derived from Principal Component Analysis (PCA), as illustrated with a refinery hydro-cracker application. In this study, our focus will be solely on holding mode operations where datasets from an actual LNG regasification terminal are utilized. Using the multivariate approach for state identification, these datasets are segmented into steady-state and transient operational modes initially, given the varying control requirements subjected to the plant's operation. To address the data's time-varying characteristics and consider the process's multivariate nature, we incorporated Dynamic Principal Component Analysis (DPCA) into the model as proposed by Srinivasan et al. (2004).
After the data is categorized into the steady and transition states, the steady states are clustered, enabling predictions for BOG in the holding mode using soft sensors, also known as virtual or inferential sensors. These sensors employ machine learning algorithms to estimate process variables in industrial settings, offering cost-effective alternatives to physical sensors while handling complex and dynamic systems. They offer a key advantage by utilizing readily available data and models to provide real-time insights, enhancing process monitoring, control, and optimization. In this study, process parameters such as tank vapor pressure, temperature, and LNG density are used to predict BOG generation.
Various models have been built in the literature to develop soft sensors for use in refinery applications. A paper by Gonzaga et al. (2009) outlines creating and implementing a soft sensor in the polyethylene terephthalate (PET) production process that provides real-time estimates of PET viscosity. Another paper by Wood (2020) developed soft sensors to determine LNG's saturated vapor pressure (SVP) for effective tank pressure control, which constantly changes due to evolving compositions. To improve operational decisions, machine learning models use readily available parameters such as density and temperature from five distinct LNG cargoes for automated SVP predictions. However, no data-driven model has been developed in the literature for BOG prediction.
This study emphasizes the need to employ a data-driven model for accurate BOG estimation, which is crucial for ensuring safety and operational efficiency, as illustrated in a case study. The estimated BOG predicted are validated by comparing them with observations from actual plant data. In summary, data-driven models are essential for BOG prediction due to their computational efficiency, capacity to capture complex relationships from extensive datasets, and real-time integration with process plants, setting them apart from analytical and numerical models.
References
Wood, D.A. (2021) Predicting saturated vapor pressure of LNG from density and temperature data with a view to improving tank pressure management, Petroleum, 7, 91-101, https://doi.org/10.1016/j.petlm.2020.04.001.
Gonzaga, J.C.B., Meleiro, L.A.C., Kiang C., Filho, R.M. (2009) ANN-based soft-sensor for real-time process monitoring and control of an industrial polymerization process, Computers & Chemical Engineering, 33, 43-49, https://doi.org/10.1016/j.compchemeng.2008.05.019S
Srinivasan, R., Wang, C., Ho, W., Lim, K. (2004). Dynamic Principal Component Analysis Based Methodology for Clustering Process States in Agile Chemical Plants. Industrial & Engineering Chemistry Research - IND ENG CHEM RES. 43, https://doi.org/10.1021/ie034051r
Srinivasan, R., Viswanathan, P., Vedam, H., & Nochur, A. (2005). A framework for managing transitions in chemical plants. Computers & Chemical Engineering. 29. 305-322, https://doi.org/10.1016/j.compchemeng.2004.09.024