Prediction of Pellet Durability Index in a commercial feed mill using multiple linear regression with variable...

Area of Science:

• Agricultural Engineering
• Animal Science
• Statistical Modeling

Background:

• Pellet quality, measured by Pellet Durability Index (PDI), is vital for feed manufacturing efficiency and animal nutrition.
• Controlling pellet quality is challenging due to numerous complex process variables.
• Existing prediction methods use either limited empirical models or complex machine learning approaches.

Purpose of the Study:

• To develop statistical regression models for predicting PDI in commercial feed manufacturing.
• To identify and describe the relationships between pellet quality and 55 potential influencing variables.

Main Methods:

• Collected a dataset of 2691 observations from a commercial feed mill.
• Transformed the PDI variable using the Box-Cox method for improved normality (tPDI).
• Developed three multiple regression models using Forward Selection, Principal Component Analysis, and Partial Least Squares, followed by variable selection.

Main Results:

• The Forward Selection model (Model 1), utilizing 9 variables, demonstrated superior performance over PCA and PLS models.
• Model 1 showed consistent prediction accuracy on training and testing data, with low error metrics (MAE, RMSPE) and a good concordance correlation coefficient.
• Expanding Temperature, Fat Content, ADF Content, and Indoor Humidity (Pelletizer) were identified as the most influential variables affecting tPDI in Model 1.

Conclusions:

• Statistical regression models, particularly the Forward Selection approach, can effectively predict pellet quality (PDI).
• Key variables such as temperature, fat, and ADF content significantly impact pellet durability.
• These models offer valuable tools for feed mills to predict and understand factors influencing pellet quality.