Predicting Porosity in Grain Compression Experiments Using Random Forest and Metaheuristic Optimization Algorithms.

Area of Science:

• Agricultural Engineering
• Machine Learning Applications
• Food Storage Science

Background:

• Long-term grain storage faces significant losses due to condensation, mold, and insect infestations, especially in large bungalow warehouses.
• Grain pile porosity is a critical factor influencing heat/moisture transfer and ventilation efficiency in stored grains.
• Accurate porosity assessment is vital for mitigating storage losses and ensuring food security.

Purpose of the Study:

• To investigate the distribution pattern of bulk grain pile porosity in bungalow warehouses.
• To develop and evaluate machine learning models for predicting grain pile porosity.
• To identify the optimal machine learning model for efficient and accurate porosity prediction.

Main Methods:

• Compression experiments were conducted to gather data for porosity prediction.
• Five machine learning models were developed: Random Forest (RF) and four hybrid models (PSO-RF, GWO-RF, SCA-RF, TSA-RF).
• Model performance was assessed using error analysis, Taylor diagrams, evaluation metrics, and multi-criteria assessments.

Main Results:

• Hybrid machine learning models significantly outperformed the standard Random Forest model.
• The Tunicate Swarm Algorithm-Random Forest (TSA-RF) model achieved the highest predictive accuracy (R²=0.9923 training, R²=0.9723 testing).
• Hierarchical prediction revealed grain porosity is higher in the center and decreases towards the edges with increasing depth.

Conclusions:

• The TSA-RF model provides a novel, efficient method for predicting grain porosity in bulk storage.
• Understanding porosity distribution aids in optimizing ventilation and preventing spoilage in grain warehouses.
• This machine learning approach enables rapid porosity assessments, contributing to improved food security.