LSTM-based control of cellulose degree of polymerization in a batch pulp digester

Abstract

Due to ever increasing demand for different types of paper, it is crucial to optimize the Kraft pulping process to achieve the required paper properties. This work proposes a framework to regulate these paper properties by building a novel closed-loop long short-term memory (LSTM)-based model predictive control system. First, a multiscale model was developed by combining the mass and thermal energy balance equations adopted from Purdue model with a layered kinetic Monte Carlo (kMC) algorithm that describes the time-evolution of microscopic events such as fiber morphology, kappa number, and cellulose Degree of Polymerization (DP). Then, this model was run over different operating conditions by varying the temperature, concentration, and cooking time to generate data. An LSTM-ANN network was trained using these datasets with a prediction accuracy of over 98% capturing the behavior of cellulose DP and considering the effects of time-varying and time-invariant operating conditions together. Finally, a closed-loop LSTM-based optimal controller was designed, which was demonstrated to achieve the target set-point values and obtain optimal constant value inputs along with time-series inputs while considering process constraints. The results showed excellent accuracy and the controller was computationally less expensive due to the use of a well-trained LSTM network in the proposed framework.