Multi-Task Regression with Process Knowledge-Based Forest Learners in Process Industries*

The multi-task regression is of great significance for engineers in process control. It can not only help achieve more than one target variable estimation at once, but more importantly, it offers an opportunity for target variables to share knowledge. However, due to the complexity and nonlinearity of the production process that contains interconnected variables in process industries, it is challenging for conventional regression methods to obtain accurate results. To fill the gaps, more black-box artificial intelligence models are proposed and applied to multi-task regression, while the process knowledge is often ignored during modeling. In this paper, a solution for adopting process knowledge-based learners for multi-task regression is proposed. Specifically, a two-stage learning structure with multiplicative integration weighted-random forest is offered. The initial estimated values of target variables are achieved from the separated base learners in the first stage, and then an integrated model is learned based on the previous predictions. In this manner, the knowledge of target variables and process knowledge among process data can ultimately be learned after training. The proposed method is finally applied to a VRM system in a case study, and the results verify its superiority over baseline methods.