Functional link hybrid artificial neural network for predicting continuous biohydrogen production in dynamic membrane...

Area of Science:

• Biotechnology
• Chemical Engineering
• Machine Learning

Background:

• Conventional machine learning methods struggle with the nonlinear dynamics of continuous biohydrogen production.
• Accurate forecasting of hydrogen production rate (HPR) and hydrogen yield (HY) is crucial for optimizing bioreactor performance.

Purpose of the Study:

• To develop and evaluate a hybrid machine learning model for predicting dynamic membrane reactor performance in biohydrogen production.
• To forecast key performance indicators: hydrogen production rate (HPR) and hydrogen yield (HY).

Main Methods:

• Developed hybrid algorithms by integrating particle swarm optimization (PSO) with functional link artificial neural networks (FLN).
• Utilized laboratory-based daily operation data with twelve input variables for model training and validation.
• Employed Shapley additive explanations (SHAP) to identify key influencing parameters.

Main Results:

• The PSO-FLN hybrid model demonstrated superior performance in predicting both HPR (R²=0.97) and HY (R²=0.80).
• Achieved low prediction errors: 0.014% for HPR and 0.023% for HY.
• Identified organic loading rate (OLR) and butyric acid concentration as key positive influencers for HPR.

Conclusions:

• The PSO-FLN model effectively handles complex, nonlinear datasets in biohydrogen production with high precision.
• This approach offers a computationally efficient method for real-time prediction of bioreactor performance.
• The findings provide valuable insights for optimizing biohydrogen production processes through advanced machine learning.