Prediction of biogas production of industrial scale anaerobic digestion plant by machine learning algorithms

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.

A data fusion system based on attenuated total reflectance mid-infrared spectroscopy and colorimetry combined with chemometrics for monitoring the fermentation process of Candida utilis

The fermentation of Candida utilis has become a cost-effective and high-yield process, characterized by its high nutritional value, high productivity, and short fermentation time. To efficiently and comprehensively monitor the fermentation process of Candida utilis, this study employed an integrated data fusion system, based on attenuated total reflectance mid-infrared (ATR-MIR) spectroscopy and colorimetry, combined with chemometrics to achieve efficient and comprehensive monitoring of the Candida utilis fermentation process. By analyzing the trends in key physicochemical indicators during fermentation, the fermentation process could be divided into four stages: 0-3 h, 3-9 h, 9-24 h, and 24-33 h. Principal component analysis (PCA) was employed to reduce the dimensionality of the highly collinear infrared spectral data to 1-10 principal components. Through the evaluation of four distinct machine learning algorithms, the optimal data fusion method and the best-performing machine learning model (random forest, with a classification accuracy of 95.30 %) were identified. PCA and linear discriminant analysis (LDA) revealed that samples from different fermentation times exhibited distinct clustering trends, although full differentiation was not achieved. To enhance classification accuracy, supervised learning models based on class labels were introduced. A comparison of classification accuracy between single signal and fusion signal approaches revealed that the optimal model achieved superior performance on the fusion dataset, with a 4-stage classification accuracy of 0.978, which was higher than that of the 11-stage classification. Quantitative prediction of key physicochemical parameters during the fermentation process was conducted using partial least squares regression (PLSR) and support vector regression (SVR) based on colorimeter dataset, ATR-MIR dataset, and fusion dataset. The data fusion strategy demonstrated excellent predictive performance for physicochemical indicators, particularly in predicting pH (R2p = 0.940, RMSEP = 0.210) and cell concentration (R2p = 0.946, RMSEP = 0.236). The color difference dataset exhibited the highest accuracy in predicting reducing sugar (R2p = 0.987, RMSEP = 0.988). The results demonstrated that, compared to PLSR, SVR demonstrated superior performance in these quantitative analysis tasks, enabling more accurate prediction of the physicochemical parameters in the Candida utilis fermentation process.

Artificial intelligence-based modeling of biogas production in a combined microbial electrolysis cell-anaerobic digestion system using artificial neural networks and adaptive neuro-fuzzy inference system

Accurate prediction of biogas production is essential for optimizing process performance, enhancing system stability, and enabling efficient resource management in bioenergy applications. The integrated microbial electrolysis cell and anaerobic digestion (MECAD) system is a novel technology that enables higher energy recovery through the application of external voltage, offering advantages over conventional anaerobic digestion (AD). Due to the energy-intensive nature of MECAD, optimizing biogas production is particularly important to ensure energy efficiency and system feasibility. In this study, the biogas production rate of a MECAD system was predicted using two machine learning (ML)-based models: an artificial neural network (ANN) and an adaptive neuro-fuzzy inference system (ANFIS). The input variables of the models-including pH, oxidation-reduction potential (ORP), total solids (TS) and volatile solids (VS) removal, hydraulic retention times (HRT), organic loading rates (OLR), applied voltage, and current production (CP)-were collected from a MECAD reactor fed with cattle manure and operated under different conditions. The ANN model achieved R2 values of 0.9844 for the testing dataset and 0.9760 for the overall dataset, with corresponding biogas production rates of 171 mL/day and 204 mL/day, respectively. The index of agreement (IA) was 0.9960 for testing and 0.9939 for overall data. Similarly, the factor of two (FA2) values were 0.9962 (testing) and 0.9956 (overall), while the mean bias (MB) was calculated as 10.05 mL/day for testing and 8.52 mL/day for overall data. In comparison, the ANFIS model yielded R2 values of 0.9811 (testing) and 0.9774 (overall), with RMSE values of 188 mL/day and 198 mL/day, respectively. The IA values were 0.9952 and 0.9943; FA2 values were 0.9962 and 0.9987; and MB values were 6.81 mL/day and 2.91 mL/day for testing and overall datasets, respectively. The results demonstrated that both machine learning-based models effectively and accurately predicted the biogas production in a laboratory-scale MECAD reactor utilizing cattle manure as the substrate.

Machine learning-based analysis of microplastic-induced changes in anaerobic digestion parameters influencing methane yield

Microplastics (MPs) present significant challenges for anaerobic digestion (AD) processes used in energy recovery from contaminated organic waste. Given that optimal AD conditions vary widely across studies when MPs are present, a robust predictive model is essential to accurately assess these complex effects. This study applied four machine learning algorithms to predict methane yield using two datasets-one with and one without MPs. Among these, gradient boosting regression demonstrated the highest prediction accuracy, with testing R2 values of 0.996 for systems without MP pollution and 0.998 with MP pollution. This model was then further optimized by removing redundant and low-importance features, refining its predictive power. Feature importance analysis revealed that digestion time and substrate organic matter content were key parameters positively correlated with methane production. In the presence of MPs, substrate pH and inoculum total solids emerged as critical factors, with partial dependence plots offering deeper insights into their optimal conditions. This research offers new perspectives on the intricate effects of MPs on methane production, which could inform the optimization of AD processes in environments contaminated by MPs.

Machine learning for enhancing prediction of biogas production and building a VFA/ALK soft sensor in full-scale dry anaerobic digestion of kitchen food waste

Based on operational data collected over 1.5 years from four full-scale dry anaerobic digesters used for kitchen food waste treatment, this study adopted eight typical machine learning algorithms to distinguish the best biogas prediction model and to develop a soft sensor based on the VFA/ALK ratio. Among all the eight tested models, the CatBoost (CB) algorithm demonstrated superior performance in terms of prediction accuracy and model fitting. Specifically, the CB model achieved a biogas production prediction accuracy (R2) ranging from 0.604 to 0.915, and a VFA/ALK R2 between 0.618 and 0.768 on the test dataset. Furthermore, the feature importance analysis revealed that biomass amount into the dry anaerobic digester was the primary factor influencing biogas production. Chemical oxygen demand (COD) and free ammonia nitrogen (FAN) were identified as the most significant factors impacting the VFA/ALK indicator during dry anaerobic digestion, collectively contributing to nearly 50% of the influence. Overall, this study verifies the feasibility of using machine learning to predict biogas production in full-scale dry anaerobic digestion and provides a crucial foundation for monitoring the stability of dry anaerobic digesters.

Ensemble machine learning prediction of anaerobic co-digestion of manure and thermally pretreated harvest residues

This study aimed to clarify the statistical accuracy assessment approaches used in recent biogas prediction studies using state-of-the-art ensemble machine learning approach according to 10-fold cross-validation in 100 repetitions. Three thermally pretreated harvest residue types (maize stover, sunflower stalk and soybean straw) and manure were anaerobically co-digested, measuring biogas and methane yield alongside eight thermal preprocessing and biomass covariates. These were the inputs to an ensemble machine learning approach for biogas and methane yield prediction, employing three feature selection approaches. The Support Vector Machine prediction with the Recursive Feature Elimination resulted in the highest prediction accuracy, achieving the coefficient of determination of 0.820 and 0.823 for biogas and methane yield prediction, respectively. This study demonstrated an extreme dependency of prediction accuracy to input dataset properties, which could only be mitigated with ensemble machine learning and strongly suggested that the split-sample approach, often used in previous studies, should be avoided.

Assessing microbial stability and predicting biogas production in full-scale thermophilic dry methane fermentation of municipal solid waste

Compared to typical anaerobic digestion processes, little is known about both sludge microbial compositions and biogas production models for full-scale dry methane fermentation treating municipal solid waste (MSW). The anaerobic sludge composed of one major hydrogenotrophic methanogen (Methanoculleus) and syntrophic acetate oxidizing bacteria (e.g., Caldicoprobacter), besides enrichment of MSW degraders such as Clostridia. The core population remained phylogenetically unchanged during the fermentation process, regardless of amounts of MSW supplied (∼35 ton/d) or biogas produced (∼12000 Nm3/d). Based on the correlations observed between feed amounts of MSW from 6 days in advance to the current day and biogas output (the strongest correlation: r = 0.77), the best multiple linear regression (MLR) model incorporating the temperature factor was developed with a good prediction for validation data (R2 = 0.975). The proposed simple MLR method with only data on the feedstock amounts will help decision-making processes to prevent low-efficient biogas production.

Machine learning assisted modelling of anaerobic digestion of waste activated sludge coupled with hydrothermal pre-treatment

This study utilizes decision-tree-based models, including Random Forest, XGBoost, artificial neural networks (ANNs), support vector machine regressors, and K nearest neighbors algorithms, to predict sludge solubilization and methane yield in hydrothermal pretreatment (HTP) coupled with anaerobic digestion (AD) processes. Analyzing two decades of published research, we find that ANN models exhibit superior fitting accuracy for solubilization prediction, while decision-tree models excel in methane yield prediction. Pretreatment temperature is identified as pivotal among various variables, and heating time surprisingly emerges as equally significant as holding time for solubilization and surpasses it for methane yield. Contrary to prior expectations, the HTP method's impact on sludge solubilization and AD performance is minimal. This study underscores data-driven models' potential as resource-efficient tools for optimizing advanced AD processes with HTP. Notably, our research spans nearly two decades of lab, pilot, and full-scale studies, offering novel insights not previously explored.

Real-time operation of municipal anaerobic digestion using an ensemble data mining framework

This study presents a novel approach for real-time operation of anaerobic digestion using an ensemble decision-making framework composed of weak learner data mining models. The framework utilises simple but practical features such as waste composition, added water and feeding volume to predict biogas yield and to generate an optimised weekly operation pattern to maximise biogas production and minimise operational costs. The effectiveness of this framework is validated through a real-world case study conducted in the UK. Comparative analysis with benchmark models demonstrates a significant improvement in prediction accuracy, increasing from the range of 50-80% with benchmark models to 91% with the proposed framework. The results also show the efficacy of the weekly operation pattern, which leads to a substantial 78% increase in biogas generation during the testing period. Moreover, the pattern contributes to a reduction of 71% in total days required for feeding and 30% in total days required for pre-feeding.

Boosting thermophilic anaerobic digestion with conductive materials: Current outlook and future prospects

Thermophilic anaerobic digestion (TAD) can provide superior process kinetics, higher methane yields, and more pathogen destruction than mesophilic anaerobic digestion (MAD). However, the broader application of TAD is still very limited, mainly due to process instabilities such as the accumulation of volatile fatty acids and ammonia inhibition in the digesters. An emerging technique to overcome the process disturbances in TAD and enhance the methane production rate is to add conductive materials (CMs) to the digester. Recent studies have revealed that CMs can promote direct interspecies electron transfer (DIET) among the microbial community, increasing the TAD performance. CMs exhibited a high potential for alleviating the accumulation of volatile fatty acids and inhibition caused by high ammonia levels. However, the types, properties, sources, and dosage of CMs can influence the process outcomes significantly, along with other process parameters such as the organic loading rates and the type of feedstocks. Therefore, it is imperative to critically review the recent research to understand the impacts of using different CMs in TAD. This review paper discusses the types and properties of CMs applied in TAD and the mechanisms of how they influence methanogenesis, digester start-up time, process disturbances, microbial community, and biogas desulfurization. The engineering challenges for industrial-scale applications and environmental risks were also discussed. Finally, critical research gaps have been identified to provide a framework for future research.

Mechanisms, methods and applications of machine learning in bio-alcohol production and utilization: A review

Bio-alcohols have been proven promising alternatives to fossil fuels. Machine learning (ML), as an analytical tool for uncovering intrinsic correlations and mining data connotations, is also becoming widely used in the field of bio-alcohols. This article reviews the mechanisms, methods, and applications of ML in the bio-alcohols field. In terms of mechanisms, we describe the workflow of ML applications, emphasizing the importance of a well-defined research problem and complete feature engineering for a robust model. Prediction and optimization are the main application scenarios. In terms of methods, we illustrate the characteristics of different ML models and analyze their applicability in the bio-alcohol field. The role of ML in the production of bio-methanol by pyrolysis and gasification, as well as in the three stages of fermentation for bioethanol production are highlighted. In terms of utilization, ML is used to optimize engine performance and reduce emissions. This review provides guidance on how to use novel ML methods in the bio-alcohol field, showing the potential of ML to streamline work in the whole biofuel field.