Multi-target deep learning models for syngas yield and exergy estimation in hybrid fixed and fluidized bed biomass-lignite gasifiers

Abstract

Hybrid biomass-lignite co-gasification presents a promising route for sustainable syngas and exergy production. This study explores the predictive modeling of gasification outputs namely CO, CO2, CH4, H2 yields and exergy values using advanced machine learning strategies. A comprehensive dataset comprising elemental compositions and reactor configurations (fixed and fluidized bed) was generated via Aspen Plus simulations. Multi-target deep learning models, including Artificial Neural Networks (ANN), Long Short-Term Memory (LSTM), Bidirectional LSTM (Bi-LSTM), and Gated Recurrent Units (GRU), were developed and evaluated through 10-fold crossvalidation and hold-out methods. All models demonstrated high predictive performance, with average R2 scores exceeding 0.97 and RMSE values remaining below 0.01 across targets. Bi-LSTM models marginally outperformed others, achieving R2 values up to 0.991. Furthermore, Van Krevelen diagrams were used to visualize the relationship between fuel composition (H/C and O/C ratios) and gasification performance across reactor types and optimization objectives. These visual diagnostics revealed distinct clusters and trends aligned with reactor behavior and compositional characteristics, offering interpretability to the model predictions. The study not only benchmarks deep learning architectures for multi-output regression in thermochemical systems but also demonstrates how visual analytics can bridge the gap between data-driven modeling and process insight.