Simulação de diferentes configurações de colunas de destilação para a separação da mistura nonilfenol-dinonilfenol com aplicação de redes neurais artificiais
Data
2022-12-22
Autores
Nunes, Julio Cesar Ribeiro [UNIFESP]
Orientadores
Falleiro, Rafael Mauricio Matricarde [UNIFESP]
Tipo
Dissertação de mestrado
Título da Revista
ISSN da Revista
Título de Volume
Resumo
O nonilfenol é um produto de grande relevância no mercado de surfactantes devido a versatilidade de aplicação em diferentes tipos de produtos químicos. A maior aplicação está presente no setor de surfactantes não-iônicos em produtos de cuidados domésticos, pessoal e agroquímicos. Para a obtenção deste produto, estudou-se diferentes configurações estruturais de colunas de destilação, a partir de simulação computacional, de um sistema de separação nonilfenol-dinonilfenol (NF-DNF), a fim de se propor a melhor configuração visualizada em questão de resultados das correntes e demanda energética do processo de separação. Além disso, a realização da modelagem matemática, através de redes neurais artificias (RNA), serviu como uma ferramenta de auxílio na tomada de decisão operacional com base em valores de entrada das variáveis de processo do sistema de separação. Através das RNAs foi possível contornar a ausência dos parâmetros de equilíbrio líquido-vapor da mistura NF-DNF, fato que torna a RNA uma alternativa de complemento na análise do estudo pela não necessidade de atender os graus de liberdade que o simulador precisa. Foram definidas as configurações estruturais a serem avaliadas a mudança dos recheios internos, com alternativa em randômico e estruturado, e alteração no diâmetro das colunas de destilação. Posteriormente foi feito o levantamento de dados técnicos da planta produtiva tomada como referência, além de resultados analíticos de correntes do processo. Foi definido o modelo termodinâmico e validado com dados de referência. Não foi identificada variação significativa nos resultados de composição das correntes de saída e da demanda energética do processo, com redução em alguns casos de até cerca de 2 % dessa demanda, positivo para o custo operacional. Para o caso da avaliação do diâmetro das seções da coluna foi visto que a condição de referência é menor do que as simuladas. Com relação a modelagem matemática, foram desenvolvidas as RNAs com o banco de dados já levantado e considerando 21 variáveis de entrada e 2 variáveis de saída. Posteriormente, foi definido o método de otimização de Levenberg-Marquardt e a tangente hiperbólica como função de ativação. O treinamento foi feito com 60 estruturas distintas em questão de número de camadas ocultas, 1 até 3, e neurônios por camada. A estrutura 21-70-20-2 apresentou o melhor resultado, ajuste e caráter de predição do modelo, além do menor erro quadrático médio, de 0,0119.
Nonylphenol is a product of great relevance in the surfactant market due to its versatility of application in different types of chemical products. The largest application is present in the non-ionic surfactant sector in household, personal care and agrochemical products. The study aims to carry out different distillation column configurations of a nonylphenol-dinonylphenol separation system, in order to provide the best configuration in terms of stream results and energy demand of the separation process. In addition, perform a mathematical modeling, through artificial neural networks (ANN), as a tool to support in operational decision making based on input values of the process variables of the separation system. Through the ANN, it is also expected to avoid the absence of the liquid-vapor equilibrium patterns of the NF-DNF mixture, a fact that makes the ANN an alternative complement in the analysis, since it does not need to meet the degrees of freedom as a process simulator needs. For the development of the study, the configurations to be evaluated on two fronts were defined, changing the column internals, with alternatives in random and structured packings, and changing the diameter of the packed distillation columns. a survey of technical data from the production plant taken as a reference was made, in addition to analytical results of process streams. Simulation started using the Aspen Plus® software and a thermodynamic model chosen and validated with reference data, to contemplate the evaluation of the configurations and visualize the most adequate one in terms of operability and cost. There was no significant variation in the results of the composition of the output streams and the energy demand of the process, with a reduction about to 2%, positive for the operating cost. For the case of evaluating the diameter of the column sections, it is seen that the reference condition is smaller than the simulated ones. Regarding mathematical modeling, artificial neural networks were developed with the database already raised and considering 21 input variables and 2 output variables. Subsequently, the Levenberg-Marquardt optimization method and the hyperbolic tangent as the activation function were defined. The training was done with 60 different structures in terms of the number of hidden layer, 1 to 3, and neurons per layer. The structure 21-70-20-2 has the best fit model and with the lowest mean squared error (MSE) when compared to the experimental data in training, validation and testing, a MSE of 0.0119.
Nonylphenol is a product of great relevance in the surfactant market due to its versatility of application in different types of chemical products. The largest application is present in the non-ionic surfactant sector in household, personal care and agrochemical products. The study aims to carry out different distillation column configurations of a nonylphenol-dinonylphenol separation system, in order to provide the best configuration in terms of stream results and energy demand of the separation process. In addition, perform a mathematical modeling, through artificial neural networks (ANN), as a tool to support in operational decision making based on input values of the process variables of the separation system. Through the ANN, it is also expected to avoid the absence of the liquid-vapor equilibrium patterns of the NF-DNF mixture, a fact that makes the ANN an alternative complement in the analysis, since it does not need to meet the degrees of freedom as a process simulator needs. For the development of the study, the configurations to be evaluated on two fronts were defined, changing the column internals, with alternatives in random and structured packings, and changing the diameter of the packed distillation columns. a survey of technical data from the production plant taken as a reference was made, in addition to analytical results of process streams. Simulation started using the Aspen Plus® software and a thermodynamic model chosen and validated with reference data, to contemplate the evaluation of the configurations and visualize the most adequate one in terms of operability and cost. There was no significant variation in the results of the composition of the output streams and the energy demand of the process, with a reduction about to 2%, positive for the operating cost. For the case of evaluating the diameter of the column sections, it is seen that the reference condition is smaller than the simulated ones. Regarding mathematical modeling, artificial neural networks were developed with the database already raised and considering 21 input variables and 2 output variables. Subsequently, the Levenberg-Marquardt optimization method and the hyperbolic tangent as the activation function were defined. The training was done with 60 different structures in terms of the number of hidden layer, 1 to 3, and neurons per layer. The structure 21-70-20-2 has the best fit model and with the lowest mean squared error (MSE) when compared to the experimental data in training, validation and testing, a MSE of 0.0119.