Application of artificial neural networks and multiple linear regression for predicting asymptotic gas production of agricultural by-products

This research explored the correlation between the chemical composition and asymptotic in vitro gas production (AGP) of diverse agricultural by-products, intending to develop predictive models for AGP using advanced computational methods. The research employed two complementary analytical approaches: artificial neural networks (ANN) and multiple linear regression (MLR), to assess their efficacy in forecasting AGP based on compositional parameters. Two datasets were utilized: a training dataset compiled from previously published literature and a testing dataset comprising experimentally derived chemical profiles and AGP measurements of selected by-products. Following the determination of chemical constituents (e.g., neutral detergent fiber [NDF], acid detergent fiber [ADF], organic matter [OM], and crude protein [CP]) and AGP values, the datasets were merged and subjected to multivariate cluster analysis. This analysis revealed two statistically distinct clusters (A and B), with intra-group similarity thresholds exceeding ۸۰% for Cluster A and ۹۰% for Cluster B. The study focused on Cluster A, which encompassed the selected by-products, for subsequent Pearson correlation and predictive modeling. Key findings included significant inverse relationships between AGP and fiber components (NDF: r=−۰.۶۵; ADF: r=−۰.۷۲), whereas positive correlations emerged with OM (r=۰.۵۸) and CP (r=۰.۴۹). Comparative model performance demonstrated ANN’s superiority (r²=۰.۷۸, RMSE=۵.۳۹) over MLR (r²=۰.۲۴, RMSE=۱۸.۳۶), highlighting its potential for accurate AGP prediction in agricultural by-products.

کلیدواژه ها:

Agricultural by-products ، Neural Networks ، Cluster analysis ، multilinear regression ، Prediction