Meta-tuning and fast optimization of machine learning models for dynamic methane prediction in anaerobic digestion

This study evaluates the performance of several optimization algorithms for tuning a data preparation and hyperparameter optimization pipeline applied to machine and deep learning models predicting methane production. Bayesian ridge regression and recurrent neural networks were applied to steady-state and dynamic datasets. Results show that 50 optimization steps are sufficient for optimal performance in simpler cases (62.8 % model accuracy). For complex scenarios, such as recurrent neural networks on dynamic datasets, extended optimization processes improve accuracy. Among the tested algorithms, Bayesian Search performed well without meta-tuning. However, meta-tuned Genetic Algorithm performed better (94.4 % vs 99.2 % baseline). Meta-tuning improves tuning parameter selection and model precision. Differential Evolution and Particle Swarm Optimization with time-varying acceleration also performed well, particularly in steady-state. These findings highlight the need to match optimization to dataset and model complexity, with meta-tuning offering advantages in challenging cases. Improved accuracy can increase revenue in flexible biogas operations.