Otimização de processos na indústria de papel e celulose: predição do número Kappa por redes neurais recorrentes

Abstract

The Kappa number is a critical parameter for fiber pulp quality control in the pulp and paper industry, directly related to residual lignin content after the Kraft cooking process. Traditional methods to determine the Kappa number, although widely employed, exhibit substantial limitations, including long response times and inaccuracies due to the oxidation of non-lignin compounds. Given this context, this study aims to develop an innovative predictive model based on Recurrent Neural Networks (RNNs) to estimate the Kappa number accurately and continuously in Kamyr continuous digesters at Klabin's industrial plant. The specific objectives include identifying relevant operational variables through Pearson correlation analysis, developing tailored prediction models for K2 and K3 pulps using LSTM and GRU architectures, evaluating model performance comparatively using quantitative metrics, and implementing the predictive solution within the industrial PI Vision™ system. For this purpose, historical operational data spanning approximately one year of industrial operation were collected and analyzed. The LSTM-based model achieved a determination coefficient (R²) of 0.8698 for K2 pulp, while the bidirectional GRU architecture model reached an R² of 0.849 for K3 pulp, demonstrating the effectiveness of RNNs in capturing the complex, nonlinear dynamics inherent to the Kraft process. Results confirm the robustness and practical applicability of the proposed methodology, enabling enhanced operational stability, optimized use of chemical inputs, and continuous improvement in pulp quality.