Estudo do processo de hidrotratamento de diesel através de simuladores comerciais e redes neurais artificiais

Abstract

Diesel plays an important role to the Brazil’s energy matrix, because it is one of the most consumed fuel at country. This study describes the hydrotreating process for petroleum refining using commercial software, PRO/II®. In addition, the artificial neural networks (ANN), using MATLAB®, were evaluated as tool for the development of virtual sensors and fault detection in chemical processes. The theoretical and practical assessment carried out in relation to the suitability of the software is justified by having acceptability in the oil and gas industry. For the oil hydrotreating associated with sulfur, the Diesel was represented by weight oil features; similar those used in refinery Abreu e Lima (RNEST). The goal of this study was to modeling in stationary condition for the reactor section the diesel hydrotreating process getting operating conditions to keep the sulfur concentration below of the ANP (Agência Nacional do Petróleo) specification, which is 10 ppm. The RNA applied to evaluate the virtual sensors was the MLP (Multi-Layer Perceptron), however, using the Levenberg-Marquardt Backpropagation (LMB) training algorithm. The coefficient of correlation was close to 1, so it means the network was satisfactory for this study. In the binary classification stage, the performance of RNA-LMB was efficient since a percentage of success above 93% was observed. In data clustering stage, used to define faults detection from parameters like flow of diesel, contaminants, pressure and temperature, the RNA-Kohonen was evaluated to check the sensitivity of the network to cluster the data with similarity. In a second stage, were inserted errors in the network, the temperature was the parameter with the greatest susceptibility to measurement errors in the industry. The lack of similarity for the input data in the RNA was well represented by self-organizing map (SOM). The simulations, as well as network training, adequately represent the analyzed processes.