Novel Intelligent System Based on Automated Machine Learning for Multiobjective Prediction and Early Warning Guidance of Biogas Performance in Industrial-Scale Garage Dry Fermentation

Abstract

Industrial-scale garage dry fermentation systems are extremely nonlinear, and traditional machine learning algorithms have low prediction accuracy. Therefore, this study presents a novel intelligent system that employs two automated machine learning (AutoML) algorithms (AutoGluon and H2O) for biogas performance prediction and Shapley additive explanation (SHAP) for interpretable analysis, along with multiobjective particle swarm optimization (MOPSO) for early warning guidance of industrial-scale garage dry fermentation. The stacked ensemble models generated by AutoGluon have the highest prediction accuracy for digester and percolate tank biogas performances. Based on the interpretable analysis, the optimal parameter combinations for the digester and percolate tank were determined in order to maximize biogas production and CH4 content. The optimal conditions for the digester involve maintaining a temperature range of 35–38 °C, implementing a daily spray time of approximately 10 min and a pressure of 1000 Pa, and utilizing a feedstock with high total solids content. Additionally, the percolate tank should be maintained at a temperature range of 35–38 °C, with a liquid level of 1500 mm, a pH range of 8.0–8.1, and a total inorganic carbon concentration greater than 13.8 g/L. The software developed based on the intelligent system was successfully validated in production for prediction and early warning, and MOPSO-recommended guidance was provided. In conclusion, the novel intelligent system described in this study could accurately predict biogas performance in industrial-scale garage dry fermentation and guide operating condition optimization, paving the way for the next generation of intelligent industrial systems.

Feature Engineering and Supervised Machine Learning to Forecast Biogas Production during Municipal Anaerobic Co-Digestion

“…While the linear regression models have simple and understandable inputs, the co-digestion process likely contained nonlinearities that were not captured by simple models. ML models have the drawback that they provide no insight into mechanistic connections between independent variables and observations; however, they are more robust to co-correlated variables, such as those found in time-series problems such as this one and have been frequently applied for biogas prediction. ,, …”

“…Higher absolute SHAP values over the entire model translate to higher model importance. Due to its ease of calculation and interpretation, SHAP values have been commonly used in other environmental studies. ,,, For our data sets, we first trained the SHAP kernel explainer on a summary of the training set based on 10 weighted k -means. For the lab data set, we then fit the explainer to the entire test set.…”

“…Another study used ML models to predict methane yield in laboratory reactors . A recent paper utilized AutoML pipelines to predict and explain biogas production from daily data from a dry fermentation system . Another study used a lab-scale system with online sensors to collect hourly data for 116 days and fit the data with deep learning models for predicting various AD parameters .…”

“… 9 A recent paper utilized AutoML pipelines to predict and explain biogas production from daily data from a dry fermentation system. 10 Another study used a lab-scale system with online sensors to collect hourly data for 116 days and fit the data with deep learning models for predicting various AD parameters. 11 Other studies that used ML models to predict biogas production were lacking a sufficient number of samples (<250) for accurate or generalizable performance, and are reviewed elsewhere.…”

Understanding Anaerobic Co-digestion of Organic Wastes through Meta-Analysis

“…Predictive models have been developed to predict the performance of AcoD systems, using data such as biomethane potential results , and operational parameters, − kinetic parameters and metabolites, as well as insights from the microbial community. , However, some of these models (e.g., artificial neural network (ANN) and random forest (RF)) are overly specific and often unrealistic due to limited and incomplete data sets. , Statistical analyses and machine learning (ML) models, including Extreme Gradient Boosting (XGBoost), Kernel ridge regression (KRR), and RF, have been used to optimize the lignocellulosic biomass (LB)-to-AM ratio for enhancing SMY . Yet, variability in the data due to different pretreatment methods, reactor designs, modes of operation, and nutrient supplementation can affect the reliability of these models.…”