Avaliação da utilização de redes neuronais aplicadas a processos quimicos

AUTOR(ES)
DATA DE PUBLICAÇÃO

2000

RESUMO

This work presents some important aspects when applying neural networks for chemicai process simulations. Neural networks are used in two different industrial processes providing a way of comparison of difficulties and potentialities for the application of this technique. Currently at Ripasa S/A. one of the largest paper industries in Brazil, the raw material classification for the pulp digester has been done using density as the property of the wood. At Ripasa S/A, the wood comes from trees of the genus Eucalyptus. The chapter 3 presents a more rigorous alternative for the criteria of raw material classification by taking into account other data such as physical properties, chemical composition of the wood and origin of the trees. A good classification should be based on the final properties of the process, which are yield. Kappa number, viscosity and residual alkalinity. Prediction of these properties by neural networks was one of the objectives of this work since experimental measurement is time-consuming and expensive. A statistical technique was used to estimate sampling and measurement errors. These values were compared with the errors generated by the neural network. Different back propagation networks were created to evaluate a suitable set of network parameters, such as number of nodes on the hidden layer, learning parameters and input range, among others. T he networks were trained by data recorded from the analysis of 165 wood samples from eight different regions. The second work (chapter 4) presents a way to predict the environmental properties of the output stream from the wastewater treatment plant at Rhodiaco Ltda, one of the major chemical plants in Brazil. The industrial plant produces purified teiephtalic acid and generates wastewater that should be treated in an activated sludge system. Back-propagation neural networks are used to predict the elimination of total organic carbon (TOO m the treatment plant, using the delta-bar-delta algorithm for estimation of weights and the sigmoid function as the neuron transfer function. The influence of input variables is analyzed, and satisfactory predicted results are obtained for some situations analyzed. The main conclusion of this work is that the neural networks can be used to establish a better operating condition for industrial chemical processes. Neural networks represent a possible aid to operations in order to predict upsets and proacltvely act to minimize output fluctuations, in the future, some work will be done to predict effluent conditions based on the actual operation data set

ASSUNTO(S)

redes neurais (computação) papel - indústria Águas residuais - purificação - tratamento biológico polpação alcalina por sulfito celulose

ACESSO AO ARTIGO

Documentos Relacionados