(432b) Developing an LSTM-ANN Model for Prediction and Optimal Control of Kappa Number and Degree of Polymerization in a Batch Pulp Digester

In the past few years, pulp products have gained attention as alternatives to non-renewables [1]. The demand for packaging paper is growing all over the world, and researchers and scientists are actively looking for new applications for wood pulp and its components [2]. But along with this growth, it is imperative to focus on high productivity and low cost. And one of the ways of cost optimization is developing efficient control strategies using modern artificial intelligence tools. Two of the main ways to achieve this is by producing pulp with desired Kappa number and degree of polymerization (DP). To attain these target pulp properties, operating conditions like cooking time, alkali concentration, and temperature should be considered [3].

Kappa number is a metric of the residual lignin content in pulps, and a lower number is desired for high-quality bright paper. The cellulose degree of polymerization is the cellulose chain length in pulp, and it plays a significant role in determining the tensile strength of end-use papers [4]. Hence, it is essential to understand their time-varying characteristics and manipulate the operating conditions to achieve a desired set-point value. In this work, a multiscale model is developed which captures the phenomena at different time and length scales in a pulp digester. A widely used mathematical model for pulping called the Purdue model and a kinetic Monte Carlo (kMC) algorithm are integrated to capture the time evolution of both Kappa number and cellulose DP [5,6]. A set of simulations were conducted by varying the operating conditions (e.g., temperature, cooking time, and NaOH concentration). To capture the trends between these manipulated inputs and the desired properties, a Long Short-Term Memory - Artificial Neural Network (LSTM-ANN) model [7,8] was trained and tested over 750 different sets of input profiles, and an accuracy of over 95% was obtained. The LSTM was used to learn very deep learning tasks requiring long-time memory using a relatively small network structure. The relation of both time-varying (temperature) and time-invariant (cooking time, NaOH concentration) operating conditions with Kappa number and Cellulose DP was captured. Modifications like early stopping, batching, and dropout were also utilized for improved prediction accuracy of neural network.

An open and closed-loop model predictive control framework was developed to obtain the optimal operating conditions utilizing the trained neural network [9]. The results showed that the set-point values for both Kappa number and cellulose DP in the pulp digester were achieved using this approach. Thus, the developed framework can be used to achieve the desired targets for the microscopic pulp properties using a well-trained neural network that determines the optimal operating conditions for both time-varying and time-invariant inputs.

References:

1. Gupta, G. K., & Shukla, P. (2020). Insights into the resources generation from pulp and paper industry wastes: challenges, perspectives and innovations. Bioresource technology, 297, 122496.
2. https://www.mckinsey.com/industries/paper-forest-products-and-packaging/...
3. Rahman, M., Avelin, A., & Kyprianidis, K. (2020). A Review on the Modeling, Control and Diagnostics of Continuous Pulp Digesters. Processes, 8(10), 1231.
4. Joutsimo, O., Wathén, R., & Tamminen, T. (2005). Effects of fiber deformations on pulp sheet properties and fiber strength. Paperi ja puu, 87(6), 392.
5. Wisnewski, P. A., Doyle III, F. J., & Kayihan, F. (1997). Fundamental continuous‐pulp‐digester model for simulation and control. AIChE Journal, 43(12),3175-3192.
6. Choi, H. K., & Kwon, J. S. I. (2019). Multiscale modeling and control of Kappa number and porosity in a batch‐type pulp digester. AIChE Journal, 65(6), e16589.
7. Hu, Y., Ni, J., & Wen, L. (2020). A hybrid deep learning approach by integrating LSTM-ANN networks with GARCH model for copper price volatility prediction. Physica A: Statistical Mechanics and its Applications, 557, 124907.
8. Ke, W., Huang, D., Yang, F., & Jiang, Y. (2017, November). Soft sensor development and applications based on LSTM in deep neural networks. In 2017 IEEE Symposium Series on Computational Intelligence (SSCI) (pp. 1-6). IEEE.
9. Terzi, E., Bonassi, F., Farina, M., & Scattolini, R. (2019). Model predictive control design for dynamical systems learned by Long Short-Term Memory Networks. arXiv preprint arXiv:1910.04024.