Modelos hibridos de colunas de destilação
AUTOR(ES)
Jones Erni Schmitz
DATA DE PUBLICAÇÃO
2002
RESUMO
Techniques like online process optimization and model predictive control require fast simulations of the process models used. However, in some cases the process models are very complex or have a high dimension. In these cases acceleration techniques that include order reduction and model simplification can be used to obtain fast enough simulations. Distillation columns are an example of process where this type of problem is usually found. In this case, besides order reduction techniques based on discrete orthogonal collocation, simplification techniques have been developed, like local thermodynamic models, based on the fact that 30-80 % of the simulation time were associated with thermodynamic calculations. In this context, the possibility arose of using neural networks to predict thermodynamic properties. It is well known that neural networks are able to represent non-linearities like those found in thermodynamic models. Neural network models can be combined with physical models, based on mass and energy balances, for example, to obtain the so called hybrid models. In this work a static hybrid model of a binary (propane/propene) distillation column has been developed, using neural networks for the calculation of equilibrium constants, of enthalpies and of their derivatives used in the simulation. When used for the simulation of the column, this model allowed for a reduction of about 55 % of the computational time, in comparison with a classical model where the thermodynamic properties were obtained using a method based on the Peng-Robinson state equation
ASSUNTO(S)
modelos matematicos metodos de simulação redes neurais (computação)
