Optimization of Cellulose Pulp Drying Using Machine Learning Techniques

This work aimed to evaluate the pulp drying process through the application of machine learning algorithms and to create a model capable of predicting the process behavior, with the goal of minimizing energy consumption. Regression techniques, including Multilayer Perceptron, Support Vector Machine, and Random Forest, were applied and evaluated using decision-making metrics for model generation. The generated model was then tested with optimization techniques such as Bayesian Optimization, Basin-Hopping, COBYLA (Constrained Optimization by Linear Approximation), L-BFGS-B (Limited Memory Broyden-Fletcher-Goldfarb-Shanno with Bound Constraints), TNC (Truncated Newton), SHGO (Simplicial Homology Global Optimization), Differential Evolution, and Trust-Region Constrained Optimization, with results compared. Two baseline cases were generated and evaluated in four distinct operational scenarios. The analysis focused on energy consumption, and the results showed potential savings of 0.671% to 1.38%, depending on the scenario considered.