Neural Network technology is an approach for describing process data behavior, using mathematical algorithms and statistical techniques. The use of neural network for modeling process is increasing in several kinds of chemical industries. This paper makes comments about successful critical factors, advantages and disadvantages of this methodology. Moreover, it presents some applications in Hydrotreating process of the petroleum refining industry. In feedstock Hydrotreating, the knowledge about characteristics of process regarding product property estimation, hydrogen chemical consumption and removal of contaminants (sulfur, nitrogen, aromatics), is very important to process optimization, product quality control and environment protection. The Neural Network technique has been used to model the behavior of the hydrogen chemical consumption, generation of light gas, the conversions of the hydrogenation of aromatic hydrocarbons (HDA), hydrodesulfurization (HDS) and hydrodenitrogenation (HDN) reactions and product physical properties. Operation conditions and some relevant feedstock properties were selected as input variables. In addition, Neural Networks have been built to predict the cetane number and stability of feedstock and hydrogenated products. The models were developed with experimental data, which were obtained in hydrogenation pilot plants from PETROBRAS. This paper presents a comparison between pilot plant data and estimated data.
The neural network technology is used to describe the behavior of systems, through mathematical algorithms and statistical techniques, which try to mimic the human brain. The purpose of developing a neural network is to produce a tool that captures essential relationships in data.
Artificial neural networks are computing tools composed of many interconnected elements called processing elements or neurons. A neuron performs a weighted summation of its input array and the application of a non-linear transfer function to this summation to give an output. The output of a neuron can be connected to input of other processing elements through weighted connections.
Learning is the process of modifying the connection weights, i.e., it is the fitting of parameters in the modeling. These weights are obtained through an optimization process with the objective of minimizing the differences between the predicted and the observed outputs. The trained network can be used to estimate unknown output data; giving to the trained network the input data of the sample not included in the data set, the network calculates corresponding output data. Although a large variety of architectures exist, the feed-forward architectures, where data propagates in only the forward direction, are more useful for steady-state modeling.
At present, there is the predominance in the li