Meta-Learning–based Prediction of Cumulative Biomethane Yield from Co-Digestion of Kitchen Waste and Chicken Manure

Accurate prediction of cumulative biomethane yield (CBY) is essential for optimizing anaerobic digestion systems treating heterogeneous organic substrates. This study develops and tests a modelling framework that uses data to predict CBY from the co-digestion of canteen kitchen waste and chicken manure, employing linear regression (LR) and an improved meta-learner. The most important operational and substrate-related variables, including pH, substrate ratio (SR), total solids (TS), and operating temperature (T), were included as model inputs. The exploratory analysis showed moderate correlations between CBY and temperature (r = 0.516) and substrate ratio (r = 0.332), indicating nonlinear and interactive effects that would be underpredicted by simple linear models. The LR model exhibited poor generalization, with training and testing R2 values of 0.269 and -0.040, respectively, indicating that it is unable to model complex process dynamics. By comparison, the meta-learner also showed much better prediction accuracy with a training R2 of 0.897 and a testing R2 of 0.725, as well as a huge reduction in error measures (test MSE = 157.177 mL2/gVS2; MAE = 9.668 mL/gVS). Additional analyses, including residual and distribution plots, were also used to demonstrate greater model robustness and lower bias. The results indicate that ensemble-based meta-learning provides a credible and explainable approach to biomethane yield prediction that could be useful for optimizing processes, selecting feedstocks, and scaling up anaerobic digestion systems using mixed organic wastes.