Prediction of syngas composition in entrained flow gasification systems using a Transformer-based deep learning model

In the gasification processes, predicting syngas composition is critical for efficient management. Machine learning (ML) models have been adopted to predict syngas, but its pilot-scale prediction has rarely been conducted. This study applies various ML models and a transformer-based deep learning into predicting for syngas composition (H2, CO, CO2, and CH4) using the O2 flow rate, coal flow rate, and O2/coal ratio. Data were developed from the gasifier of bituminous coal using at a pilot-scale entrained flow gasification. Syngas quality indicators (H2/CO and H2+CO) were additionally predicted. The three ML models used in this study are Random Forest (RF), Gradient Boosting Machine (GBM), and extreme Gradient Boosting (XGBoost), and the Feature Tokenizer-Transformer (FT-Transformer) was used as a deep learning model. Their performances were measured using the coefficient of determination (R2) and the mean squared error (MSE). Shapley Additive Explanations (SHAP) were used to identify the key input variables. Our results showed that the largest test R2 and smallest test MSE were observed with FT-Transformer for CO (0861, 0.229), CH4 (0.891, 0.026), and H2/CO (0.893, 0.0), and with RF for H2(0.824, 0.093), CO2 (0.854, 0.502), and H2+CO (0.827, 0.533). While RF exhibited better performance for certain prediction targets, it demonstrated consistent overfitting issues across syngas composition. In contrast, FT-Transformer showed decent performance without overfitting issues. The SHAP analysis showed the O2/Coal ratio overall had the greatest impact on the prediction of syngas composition. This study demonstrated the applicability of FT-Transformer for predicting the syngas composition. The findings of this study would contribute to the efficient management of various gasification processes.