Numerical modeling of high-pressure phase-equilibria data using neural networks
Date
2020-12
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Stellenbosch : Stellenbosch University
Abstract
ENGLISH ABSTRACT: The design of process systems is greatly dependent on phase behaviour data, which can be predicted using equations of state(EOSs). These models, however, often fail to predict the behaviour near the mixture critical region. A more accurate and reliable method for predicting thermodynamic behaviour in the vicinity of the mixture critical region is therefore required. The aim of this project was to model the vapour-liquid equilibrium of binary systems containing supercritical CO2 and hydrocarbons using Artificial Neural Networks (ANNs). The bubble and dew point pressures were predicted as a function of functional group, a centric factor, critical temperature and pressure of the hydrocarbon, system temperature and CO2composition of the liquid and vapour phases.The ability of ANNs to predict the vapour-liquid phase equilibrium of binary systems was evaluated by modelling different systems and comparing the results to experimental data and EOS models. Case study 1 considered binary systems containing only CO2and alkanes. Case study 2 considered binary systems of CO2and various hydrocarbons, increasing the complexity by adding various functional groups. The hydrocarbons included alkanes, alcohols and carboxylic acids.
The modelled results from case study 1 and 2 showed that the phase equilibrium of both simple and complex binary systems can be modelled using ANNs. After investigating the structure of the neural networks, the chain length and critical pressure of the hydrocarbon were eliminated as input parameters for case studies 1 and 2. The system temperature and liquid and vapour compositions of CO2were relatively more important compared to other input parameters for case study 1 where the critical and system temperatures and CO2composition of the vapour phase had a higher relative importance for case study 2. Using a feed forward neural network with two hidden layers and the log-sigmoid transfer function resulted in the optimum results for both these studies. Case study 1 and 2 resulted in acceptable 𝑅2and 𝐴𝐴𝐷%values for the training and testing data over the entire range. 𝑅2was 0.992 and 0.991 for case study 1 and 0.949 and 0.995 for case study 2 for the training and testing data sets. 𝐴𝐴𝐷%was 9.7% and 5.6% for case study 1 and 16.4% and 7.1% for case study 2 for the training and testing data sets. The ANN models were able to predict the phase behaviour over the entire range of compositions including the mixture critical region, whereas the EOS correlation models (the RK-Aspen EOS)failed to converge in the mixture critical region.
Case study 3 considered the optimisation of ANNs as used in published articles by using the methodology and outcomes as used and concluded in case studies 1 and 2. The main difference in the methodology was the way the validation and test sets were divided: these sets consisted of complete binary systems instead of single data points extracted from binary systems. Although worse results
were obtained using this methodology, the results were still acceptable. Using two hidden layers improved the accuracy of the results obtained by case study 3.
AFRIKAANSE OPSOMMING: Raadpleeg teks vir opsomming
AFRIKAANSE OPSOMMING: Raadpleeg teks vir opsomming
Description
Thesis (MEng)--Stellenbosch University, 2020.
Keywords
High pressure, UCTD, Equilibrium, Liquid-liquid, Vapors, Binary vapor systems, Artificial Neural Networks