Machine Learning and Derivative-Free Optimization for PID Tuning: Case Study of Improved Black Liquor Concentration Control

Proportional-Integral-Derivative (PID) controllers are instrumental in managing industrial processes. Their effectiveness hinges on the precision of their tuned parameters. Consequently, it becomes essential to monitor their performance frequently and re-tune them regularly to ensure optimal operation. However, conventional tuning methods require significant effort and expertise. This research tackles this challenge by using historical data to construct machine learning (ML) models, coupled with optimization algorithms, to streamline PID tuning. Specifically, we propose integrating an Explainable Boosting Machine (EBM) as an ML model and harnessing Bayesian Optimization (BO) within a comprehensive PID tuning framework. EBM stands out for its ease of construction and accuracy. The synergistic combination of EBM and BO yields an effective solution, as demonstrated through a case study involving black liquor concentration control in a multiple-effect evaporator system within kraft pulp manufacturing.