Linear and Non-Linear Soft Sensors for Predicting the Research Octane Number (RON) through Integrated...

Area of Science:

• Chemical Engineering
• Process Control
• Machine Learning

Background:

• Research Octane Number (RON) is critical for gasoline quality but is measured offline, causing delays and hindering process control.
• Current standard methods for RON determination are time-consuming, expensive, and provide infrequent data.
• Challenges in process control arise from delayed and sparse RON measurements.

Purpose of the Study:

• To develop and validate inferential models for predicting RON using real-time process data from a catalytic reforming unit.
• To compare the performance of various linear and non-linear machine learning models for RON prediction.
• To address data complexities such as outliers, missing values, and multirate/multiresolution data.

Main Methods:

• Development of 20 predictive models, including linear and non-linear machine learning approaches.
• Utilized a robust Monte Carlo double cross-validation for model assessment and comparison.
• Incorporated data preprocessing steps to handle outliers, missing data, and process dynamics.

Main Results:

• Achieved low Root Mean Square Error (RMSE) values close to 0.5 under testing conditions.
• Penalized regression methods and Partial Least Squares (PLS) demonstrated superior performance.
• Successfully handled multirate and multiresolution data, outliers, and missing values.

Conclusions:

• The developed inferential models enable accurate and timely prediction of RON.
• Improved process management and operational efficiency through real-time RON insights.
• Facilitates effective use of heating utilities, leading to reduced costs and emissions.