MSP-LLM: A Knowledge-Guided Multi-Scale Time-Series Language Model for Slag Tin Forecasting in Top-Blown Furnace Smelting

Slag tin content is a key quality indicator in top-blown furnace tin smelting. For accurate forecasting, noisy, minute-level process signals must be linked to sparse and delayed slag assays. Limited research has been conducted on the use of large language models (LLMs) for this mixed-frequency metallurgical soft-sensing task. To address this problem, a Metallurgical Semantic Prototype-guided Large Language Model (MSP-LLM) is developed for slag tin forecasting. Each 6 h furnace window is represented by temporal patches of 10, 30, and 60 min and reprogrammed into a frozen DistilGPT2 embedding space through metallurgical semantic prototypes. Furnace-state context is supplied by a structured process-summary token. Experiments are conducted using industrial process logs and 1472 slag assay labels collected from July to December 2025, with 295 labels reserved for chronological testing. Across five training seeds, MSP-LLM achieves the best overall performance among the evaluated baselines, with mean MAE, RMSE, and R2 values of 0.901, 1.136, and 0.773, respectively. Component analysis shows that multi-scale temporal patching and semantic prototype-guided reprogramming jointly improve process-aware token representation. Experimental results show that MSP-LLM is effective for metallurgical soft sensing with sparse assay targets.