Proactive Prediction of Total Volatile Fatty Acids Concentration in Multiple Full-Scale Food Waste Anaerobic Digestion Systems Using Substrate Characteristics with Machine Learning and Feature Analysis

Volatile fatty acids (VFAs) are one of the key intermediates that easily accumulate in food waste (FW) anaerobic digestion (AD) system. Therefore, VFAs prediction has been suggested for proactive control system. Total VFAs concentration (TVFA) in full-scale FW AD reactors was predicted using substrate characteristics and alkalinity of each reactor with machine learning approach. Data collected from ten different full-scale AD sites showed highly skewed distribution, which is a common problem in real situations and difficult to obtain satisfying model performance. Therefore, Box-Cox transformation was used to transform the distribution of the target variable and proved the improvement. Among seven candidate models, XGBoost was selected and feature selection based on permutation importance selected intermediate alkalinity to partial alkalinity ratio and acetate concentration of FW as the best out of 40 features. The evaluation was based on cross-validation (CV) with R 2 score metric. As a result, the model showed R 2 -CV of 0.64 and 0.51 on train and test data, respectively. The model was interpreted with shapley additive explanations (SHAP) to examine the local contribution of each feature and the relation between variables. The established model is beneficial to briefly predict [TVFA] in reactors’ effluent using substrate characteristics. Graphical abstract