Continual Learning-Based Neural Network With Time Embedding for Silicon Content Prediction in Blast Furnace

The prediction of molten iron silicon content ([Si]) is crucial for blast furnace operation. Nowadays, neural networks are emerging as the most advanced model for [Si] prediction tasks. However, the time-varying nature of the blast furnace system makes it difficult for neural networks to maintain long-term prediction accuracy. Besides, the practical implementation of the neural network model counters the problem of irregularly sampled data due to manual intervention. To solve the above issue, we proposed a novel continual learning based hybrid update scheme for our specially designed neural network model to perform long-term prediction tasks, together with a residual time-embedding layer to model the irregularly sampled time series. The hybrid update mechanism improves the plasticity of individual model components, thereby enhancing the model’s capability to track changes in system dynamics over time. Among these components, the time embedding layer transforms irregular timestamp information into a frequency-domain representation, enabling effective alignment and similarity computation. This design improves the model’s ability to process irregularly sampled time series data. Finally, we validate our proposed method with 4-year operation data from #2 blast furnace of Liuzhou Steel in China. Experiments show that only our method can maintain high prediction accuracy for the 4-year range while facing regularly sampled raw industrial.