Machine Learning Model for the Evaluation of Biomethane Potential Based on the Biochemical Composition of Biomass

Abstract

Biomethane potential (BMP) is often used to evaluate the biogas potential during anaerobic digestion (AD). However, BMP tests are complex and time-consuming. Therefore, the present study presents a hybrid model in machine learning (ML) for evaluating and predicting BMP based on biomass biochemical composition. Generative adversarial network (GAN) is combined with different ML models to model and predict BMP for different biomass materials. The models were trained on 64 experimental datasets (original datasets) and a combination of a GAN and the original datasets (augmented datasets). The gradient boost regression (GBR) model performed very well on the training set with both datasets compared to the support vector machine (SVM), Artificial neural network (ANN), and random forest (RF). RF and GBR models performed very well when trained with the combined GAN and original datasets, with RF slightly outperforming GBR on the test datasets (R2 score of 0.9106 vs. 0.9177). This indicates that the models benefit from the additional data generated by the GAN. The GBR model trained with the GAN and original datasets combined outperformed the RF model on the test set, with an R2 score of 0.9177 vs. 0.9106. A comparison between three different hyperparameter optimization methods (grid search, particle swarm optimization, and Bayesian optimization) showed that the grid search optimized model offers a balanced performance with an R2 score of 0.9994 and a marginal improvement on the test set with an R2 of 0.9213. Feature analysis results demonstrate that cellulose has the most influence on BMP.