Molten steel quality control in LF refining furnace based on Hybrid Neuro-Symbolic dynamic adjustment

Current prediction models for LF refining based on big data algorithms have achieved notable improvements in prediction accuracy. However, most existing models lack real-time operational decision support and effective integration with industrial control systems, preventing their deployment for closed-loop optimization in practical steelmaking environments. To address these limitations, this study proposes a dynamic quality control approach based on the Case-Based Reasoning (CBR) and Attention-augmented Long Short-Term Memory (Ata-LSTM) hybrid model. In this framework, CBR is employed to retrieve and optimize historical cases for process guidance, while LSTM is utilized to capture temporal dependencies and enhance predictive performance. Furthermore, an attention mechanism is introduced to adaptively emphasize metallurgically critical time steps. The combination effectively improves both the accuracy and interpretability of endpoint predictions for molten steel composition and temperature. Industrial application results demonstrate that the model achieves effective process optimization in terms of refining operation time and slag consumption. The proposed model achieves an accuracy rate of 90.48% for endpoint C content within a ±0.01% error range, 88.89% for Si within ±0.01%, 88.10% for Mn within ±0.01%, and 94.44% for S within ±0.005%. Additionally, the control accuracy for endpoint temperature remains within ±7 ℃ and ±5 ℃, with accuracy rates of 96.03% and 88.89%, respectively. These results verify that the proposed Hybrid Neuro-Symbolic model provides strong support for the intelligent, low-carbon, and sustainable development of LF refining processes.