Artificial Neural Network For Predictive Modeling of Quinoline Insolubles in Pitch in A Steel Industry: Quality and Sustainability

Objective: The objective of the present work was the development of an intelligent system based on artificial neural networks, supervised and with batch learning, to predict the quinoline insoluble quality (IQ) parameter of tar, using a set of data from a steel industry. Theoretical Framework: Tar, derived from coal tar, is widely used in the production of anodes for the aluminum industry, and its quinoline insoluble (IQ) content is a crucial parameter, influenced by factors such as coal mixing, coking parameters and heat treatment. IQ forecasting is challenging due to the complexity of the processes involved. To address this question, artificial neural networks (ANNs) have been applied to predict the IQ of tar, taking advantage of its ability to handle complex, nonlinear systems. Method: In this study, a feedforward RNA model was used, trained with the Levenberg-Marquardt algorithm. The model has been configured with two hidden layers and the tansig activation function. The training was carried out with real data from polymerization reactors, including variables such as level, temperature, nitrogen rate, feed flow rate of the polymerizers and tar IQ. The network was configured with 2000 interactions and 20 neurons in the hidden layer, obtaining a training error (MSE) of 1x10-20. Results and Discussion: The analysis of the results revealed that temperature and nitrogen flow have the greatest influence on the IQ of the pitch, with higher temperatures and higher flow rates resulting in a higher IQ. Research Implications: ANN modeling is effective in predicting and optimizing industrial processes, especially in the analysis of large volumes of data and complex systems, contributing to quality control in pitch production and the improvement of industrial processes. Originality/Value: The main contribution of this work lies in the demonstration of the ability of artificial neural networks to accurately model the content of Quinoline Insolubles (IQ) of the pitch, even in the face of the significant dispersion observed in the input data. This finding is particularly relevant because the high variability of the data often poses a challenge for traditional predictive models. The precision achieved by the neural network, therefore, proves the feasibility and potential of this approach to overcome the limitations inherent to the complexity of the pitch production process and optimize its control, opening new perspectives for the steel industry