A second decomposition-based CNN-LSTM model for predicting sintered ore alkalinity

Abstract

Stable control of sintered ore alkalinity is of great significance for improving furnace efficiency, reducing energy consumption, and minimising carbon emissions. Furthermore, advanced prediction of sintered ore alkalinity is key to achieving stable control. This study innovatively proposes a VMD-EEMD-CNN-LSTM hybrid model that combines the dual signal processing of Variational Mode Decomposition (VMD) and Ensemble Empirical Mode Decomposition (EEMD), along with the local feature extraction of Convolutional Neural Networks (CNN) and the temporal modelling advantages of Long Short-Term Memory networks (LSTM), to achieve multi-scale and precise prediction of sinter bed alkalinity. Validation based on industrial data from a steel plant over five consecutive months (1037 samples) shows that after VMD-EEMD decomposition, the CNN-LSTM model's prediction performance is significantly improved, with the coefficient of determination (R²) increasing from 0.73 to 0.97, and the Root Mean Square Error (RMSE) decreasing from 0.017 to 0.0071. Comparative experiments show that under the same decomposition conditions, the proposed model's prediction accuracy (R² = 0.97) outperforms CNN (0.95), LSTM (0.95), and BP neural networks (0.954). In the multi-step prediction task, the first-step prediction of the model achieves an R² of 0.97, and the third-step prediction RMSE is 0.849, demonstrating better stability compared to VMD-EEMD-CNN (0.83), VMD-EEMD-LSTM (0.78), VMD-EEMD-RNN (0.84), VMD-EEMD-BP (0.79), and VMD-EEMD-SVM (0.82). The research findings provide a reliable theoretical foundation and technical pathway for intelligent control of the sintering process.