Track: Data Analytics

Abstract

Gas condensate recovered from the natural gas is a valuable liquid product. It must be processed or treated to a commercially acceptable form for its storage or export. The treatment of gas condensate generally involves the separation of any dissolved gases such as light hydrocarbon components (i.e., methane and ethane) along with decreasing the acid contents (i.e., H₂S, mercaptans, etc.), water and salt contents to the desired standards in order to make it environmentally safe for storage and transportation purposes. The process of stripping the light end components from the heavier hydrocarbons is referred as condensate stabilization which is primarily performed to reduce the vapor pressure of condensate liquids to avoid the generation of vapor phase when transferring them to atmospheric tanks. Condensate stabilization can be achieved by either flash vaporization or fractionation. However, condensate stabilization by fractionation is a popular choice in industry as it has the capability to produce condensate liquids of desired vapor pressure in a single tower process. It is a complicated chemical process, thus requires significant efforts to develop accurate and reliable models with the objective to minimize the operational costs. Lately, the modeling approach that deals with input-output operating data has gained significant attention. Machine learning has been emerged as a proven and alternative modeling approach that based on operating data, which is readily available.

This work therefore employs artificial neural network (ANN) in order to build a model that can predict the performance of the condensate stabilizer unit. A large dataset of an industrial condensate stabilizer, comprising of operating data of input-output variables, is used to evaluate the developed model. To develop the ANN model, inlet gas flowrate, inlet gas temperature, column temperature, column pressure, condensate flowrate, condensate temperature, reboiler temperature and steam flowrate are implemented as model inputs, whereas RVP, H₂S contents and water contents are used as network outputs. The dataset is split randomly into two subsets for the purpose of training and testing the model. The training subset of data is employed to train the network by developing the network weights whereas the testing subset is used to estimate and compare the output of developed model against the independent operating data. The main objective of developing this ANN model is to predict the important parameters of the final stabilized liquid, which are RVP, H₂S contents and water contents. The predictions from the developed model are compared with the testing subset of operating data. To evaluate the model performance, mean squared error (MSE) and coefficient of determination (R²) are investigated. The results obtained from this case study show that the developed model has the potential to offer reliable and accurate predictions. As future work, an integrated framework will be developed where data-driven surrogate models will be integrated into optimization framework to find the optimal values of the variables corresponding to the minimal operational costs or energy demands. The integrated framework will be able to help the gas industry to simultaneously achieve the process efficiency, profitability and safety.