Innovative Approach to Characterize Cheese Whey Anaerobic Digestion Using Combined Mechanistic and Machine Learning Models

Abstract

Whey, a cheese production byproduct, can be anaerobically digested to reduce pollution and generate energy. Yet, stability is challenging due to organic content sensitivity and influent fluctuations. The present work attempts to implement the mechanistic model and machine learning (ML) models (support vector regression (SVR) and artificial neural networks (ANNs)) together to predict the concentration dataset of substrate 1 (S₁) (i.e., carbohydrates and proteins), substrate 2 (S₂) (i.e., glucids and amino acids), VFA, and methane (CH₄) as a function of input independent variables, namely time and organic loading rate (OLR). The R² values for S₁, S₂, VFA, and CH₄ obtained through the mechanistic model remained as 0.953, 0.918, 0.84, and 0.976, respectively; for ANN models, 0.982, 0.928, 0.958, and 0.99; and for SVR models, 0.984, 0.939, 0.938, and 0.999, respectively. ML models have been discovered to be among the most precise and versatile compared to the mechanistic model. Moreover, other performance metrics, such as RMSE (0.022–2.177), MRE (0.007–0.100), and AARE (0.008–0.104) for ANN and RMSE (0.083–1.961), MRE (0.021–0.091), and AARE (0.037–0.089) for SVR, are obtained, indicating good prediction performances for both ML models. SVR and ANN models excel, aligning concentration curves to the optimum line when input parameters are adjusted, unlike the subpar traditional-based mechanistic model. Therefore, ML methods offer a tool to predict anaerobic digestion more effectively, enhancing design and operations.

Graphical Abstract