Hydrogen production prediction from Co-gasification of biomass and plastics using attention-gated MLP model

Research explores combining plastics and biowastes to decrease carbon footprints in the cement, iron, and steel industries. Co-gasification of these materials cuts carbon and converts it into hydrogen and syngas. The escalating demand for alternative energy sources has driven significant interest in hydrogen as a clean and efficient fuel. However, accurate prediction techniques are indispensable for ensuring the viability and sustainability of hydrogen production processes. While artificial intelligence models have shown promise in various fields, including computer vision and hydrogen prediction, they often encounter limitations such as low prediction accuracy with few data samples, lack of interpretability in yield result explanations, and susceptibility to adversarial attacks. Addressing these challenges is critical for advancing the reliability and applicability of hydrogen production prediction models. This study presents an attention-gated MLP model (agMLP). This novel end-to-end approach integrates MLP and attention mechanisms to enhance the accuracy and interpretability of hydrogen production prediction from the co-gasification of biomass and plastics. Furthermore, several explainable AI models were implemented to explain the proposed agMLP prediction for human understanding while using the Williams plot to indicate the applicability of the proposed agMLP. Further analysis was carried out using the Fast Gradient Sign Method and Projected Gradient Descent adversarial attack on the proposed agMLP model based on 0.2, 0.4, 0.6, and 0.8 perturbation to establish the associated risk of artificial intelligence model prediction via adversarial attack. Lastly, to compare the result of the agMLP model, nine (9) machine learning regression models and three ensemble models were developed and evaluated. The implemented model's performance was assessed using various evaluation metrics, including Mean Absolute, Mean Squared, Root Mean Squared, Root Mean Squared Logarithmic Error, and R2. The agMLP outperformed existing models' prediction accuracy while providing transparent and interpretable insights into the underlying factors influencing hydrogen production with an MAE of 0.061, MSE of 0.038, RMSE of 0.195, R2 of 0.997, MSLE of 0.000, and RMSLE of 0.005. Results from the explainable models show that independent variables with the highest relative relevance scores are: Percentage of plastics > HDPE particle size > RSS particle size > Temperature, while in comparison to size, temperatures between 500 °C and 900 °C have less of an effect on H2 generation. The agMLP model is subjected to FGSM and PGD adversarial attacks, resulting in significant performance degradation. The result of this study not only illustrates the potential of deep learning models in predicting hydrogen production and demonstrates the implications for optimizing hydrogen production processes, facilitating informed decision-making, and accelerating the transition towards sustainable energy systems.