GAN-MAML strategy for biomass energy production: Overcoming small dataset limitations

Data-driven machine learning (ML) has the potential to improve biomass energy production methods such as incineration, composting, pyrolysis, and anaerobic digestion. However, due to the scarcity and variability of data in the field, there is currently no universal model that excels across all production technique domains. To address these challenges, this study combines Model-Agnostic Meta-Learning (MAML) with Generative Adversarial Networks (GANs) to improve ML generalization in complex biomass conversion scenarios. Compared to the best ML models, the GAN-MAML models demonstrated superior performance in various domains and scales. During the testing phase, the GAN-MAML models mitigated the limitations associated with data scarcity and variability, improving performance by up to 33.1 % over the best ML models. This represents a significant improvement over the initial increase of up to 28.2 % for the MAML models. Subsequently, models trained on literature data were deployed in a real energy production factory and predicted samples they had never seen before. The results showed that the GAN-MAML models outperformed the best ML models, with the highest improvement being 28.6 %. This is a significant improvement over traditional ML and offers a flexible framework for research and practice in biomass energy production, promoting sustainable environmental solutions.