Search

Browse

• Communities
• Authors
• Year
• Titles
• Subjects

My Space

• Login

Neural network modeling unveiling the relationship between hydrothermal pretreatment and anaerobic digestion of biomass

Mouzourakis Andreas

Full record

---

Summary

The utilization of lignocellulosic biomass is an effective strategy for reducing carbon emissions, aligning with modern requirements for environmental sustainability and energy conservation. Anaerobic digestion is an energy-efficient technology for mitigating greenhouse gas emissions and recovering biofuels. The hydrothermal pretreatment assisted anaerobic digestion of biomass represents a method to enhance the cumulative methane yield, while the efficiency of the process depends on various physicochemical factors. The purpose of this study was to develop an artificial neural network (ANN) model to predict the cumulative methane yield of lignocellulosic biomass produced through anaerobic digestion. Data was collected from 8 publications and 42 different hydrothermal pretreatment conditions. The model included 12 input parameters covering the physicochemical properties of the biomass, the pretreatment conditions and the conditions of anaerobic digestion. The output of the model was the daily accumulated methane production.The model was built effectively running 65 tests, changing each time the number of hidden layers and neurons, using evaluation metrics as a yardstick. The best performing model consisted of 2 hidden layers with 14 neurons each and achieved a fitting accuracy of R2 = 0.9364. The mean squared error (MSE) for the model was 1.4×10-4, while the same metric for training, validation and testing was 4.7×10-4, 4.9×10-3, and 5.6×10-3, respectively. Blind tests were then applied with 8 different datasets that the model had never run before to observe its prediction capabilities. For each blind test, a comparison was made between the experimental and predicted data. The results from the blind tests showed that 6 trials achieved a fitting accuracy higher than R2=0.90 confirming the consistent performance of the model. In contrast, 2 trials exhibited lower performance with accuracies of R2=0.7854 and R2=0.4409, respectively.Furthermore, a decision tree analysis was performed with SPSS software to examine the significance of feature parameters. Subsequently, utilizing the analysis, the two least significant input features were removed, and two models with 11 and 10 features were developed, maintaining the same architecture as the original model. The purpose of this process was to determine the impact that the removal of these features had on the model's performance.The developed model will be further used for rapidly predicting the production of methane by analyzing only the physicochemical parameters, pretreatment conditions, and anaerobic digestion conditions, thus alleviating the need for expensive experimental procedures. This approach saves time and effort, which would otherwise require a significant period, thereby contributing to the optimization of the process and the technoeconomic assessment of the technology.

Available Files

Services