Predicting the mean low compressive strength of pellets in a straight grate furnace

Abstract

Compressive strength of pellets is a critical indicator of pellet physical performance, and low-strength pellets are particularly detrimental to the smooth operation of blast-furnace smelting. Traditional strength testing methods suffer from low sampling frequency and data latency. To address these issues, the present study proposes a stacking-based feature-fusion model for predicting the mean low compressive strength (MLCS), leveraging BO-XGBoost and BiGRU-Attention. First, shallow features with clear physical significance – such as average temperature, high-temperature residence-time ratio and oxidation rate – are extracted from the multi-physical field state of the layer and combined with process parameter features to constitute the ‘shallow feature’ set. Simultaneously, a pre-trained convolutional autoencoder (CAE) is employed to extract deep features characterising the spatial distribution within the multi-physical field data. Building upon these, a stacking model is constructed: a BO-XGBoost submodel for shallow features and a BiGRU-Attention submodel for deep features serve as base learners, while linear regression functions as the meta-learner to integrate both feature types effectively. Experimental results demonstrate that the model utilising deep features achieves higher predictive accuracy than that using only shallow features, indicating that deep features encapsulate richer information related to compressive strength. The proposed model attains a maximum R² of 0.82, offering a novel approach for predicting the MLCS values. Ultimately, this method can serve as a proactive guidance tool, enabling timely process adjustments to prevent quality deviations in pelletising plants.