Machine learning based modeling and optimization of an industrial thermal cracking furnace

Abstract

Machine learning methods can capture the distinctive characteristics of a system without any prior knowledge of the process given enough actual data. In addition, they are well suited to represent systems that are complicated for first principles modeling and have many unmeasured disturbances. Accordingly, data-based modeling for the thermal cracking furnace is a promising study using actual process data set and various machine learning methods. The study's focus is on the machine learning prediction of time-series Controlled Variables (CV), which is a prerequisite for using an Advanced Process Control (APC) system in a petrochemical plant. The most crucial component of an APC system is the prediction of the controlled variables and the adjustment of those anticipated values to bring them within the user's chosen range (Lee et al., 2023). Predicting the controlled variables is our main goal in this investigation. We specifically used a variety of machine learning approaches to forecast future controlled variables by utilizing historical controlled variables. In this study, the cycle time of the furnace of a visbreaker unit and the temperature of the hottest zone of the furnace are modeled using different machine learning methods such as Support Vector Machines, Multiple Linear Regression, Decision Tree, Random Forest, and Artificial Neural Networks. Although the Random Forest model is good at predicting temperature and remaining day for the shut-down time, ANN model is used for process optimization purposes, by incorporating it in the fitness function of the genetic algorithm. When using a genetic algorithm (GA) to optimize a model for a specific task, the choice of the model as the fitness function is crucial. The fitness function evaluates how well a particular solution (set of model parameters or hyperparameters) performs the task at hand. The reason for using an Artificial Neural Network (ANN) as a fitness function in a genetic algorithm instead of a Random Forest (RF) is search space and differentiable nature of the ANN structure. Having estimated the cycle time by training the machine learning models, the inverse problem is attempted to solve such as calculating the optimal values of the features (controlled variables) for maximizing the operation time of the process within certain limits. This optimization problem is solved by sampling-based optimization methods formulating the trained machine learning models as fitness function. In this way, the necessary manipulated variables will be adjusted by the controller so that the unit can operate in the most efficient way.