Análise de emissões em caldeiras de recuperação química de fábricas de celulose Kraft: predição e análise de sensibilidade com redes neurais artificiais

Abstract

Globally, there is a movement of greater social-environmental awareness, based on the interests of society, institutions and governments. As a result, there is an increasing pressure on industrial sectors for the development of cleaner and more sustainable processes. One of the concerns is related to emissions of gases, among them, carbon oxides (CO and CO2), nitrogen (NOx) and sulfur (SOx), in addition to dust emissions, suspended particulate matter, especially when industries are located close to cities. In this scenario, the monitoring and control of emissions gains importance, and alternative methodologies to traditional, usually costly, methods have been developed. Among the solutions in process monitoring, stand out the virtual sensors based directly on historical data from the operations. In this context, the objective of the present work was to build virtual sensors for gas and particulate material emissions monitoring from equipment in pulp mills and then analyse the variables that present the great influence over those emissions. The first case involves a chemical recovery boiler of a pulp industry in Brazil. A virtual sensor was constructed for the estimation of sulfur oxide (SO2) emissions. And the second case study refers to the chemical recovery boiler in a cellulose industry in Finland, with a focus on emissions of particulate matter. For both applications, the virtual sensors are based on a neural network with MLP (Multi-Layer Perceptrons) architecture, the most usual in chemical engineering applications in general. In order to verify their performance, the results were compared with multiple linear regression model. After varying a set of parameters, such as type of activation function and number of hidden neurons, the neural network that best fitted the data reached a linear correlation coefficient (r) higher than that of linear regression multiple, for both case studies, r_mlr = 0:764 e r_mlp = 0:939, for the boiler in Brazil, and r_mlr = 0:6974 and r_mlp = 0:86, for the boiler in Finland. In the end, the process variables that most influence the results of the virtual sensors, for each case, were identified by applying a sensitivity analysis technique. For the first case study, the most influential variables were tertiary air flow, primary air flow, black liquor flow and primary air temperature. For the second case study, the most influential variables were secondary air flow, black liquor flow and its solids content, diluted non-condensable gas flow and fuel oil flow.