High-Accuracy Prediction of Pelletized Ore Performance Using an Integrated LightGBM-XGBoost Chain Model

The compressive strength and low-temperature reduction pulverization index of iron ore pellets are critical for smelting efficiency and carbon emissions. Existing predictive models often lack robustness and fail to capture the complete "process-structure-performance" relationship. This study proposes a chained prediction framework to address this gap. To overcome the limitations of a small experimental dataset (N = 29), a Generative Adversarial Network (GAN) was used for data augmentation. The utility of the synthetic data was validated not only by statistical similarity using Kolmogorov–Smirnov tests but also through a rigorous downstream validation experiment on a hold-out set of real data, confirming its ability to enhance model generalization. A heterogeneous LightGBM-XGBoost chain model was developed to first predict microstructural parameters from process conditions and then predict macroscopic performance. A quantitative analysis of error propagation was conducted to assess the robustness of the chained architecture. Results show the integrated model’s predictions are statistically significant, reducing global MAE and RMSE by 71.97 and 65.48 pct respectively, and increasing \({R}^{2}\) to 0.895 compared to a baseline model. By explicitly modeling the process-structure-performance linkage and incorporating robust validation, this work provides a data-driven tool for optimizing pellet production for low-carbon, intelligent smelting.