Prediction of Key State Parameters in Blast Furnace Ironmaking: Mechanism–Data Fusion and Cross-Condition Generalization

Abstract

Accurate prediction of blast furnace state parameters is complicated by partial observability, heterogeneous sampling frequencies, delayed laboratory measurements, and changes in operating conditions. This critical review synthesizes recent journal and conference literature on the prediction of hot-metal silicon content, hot-metal temperature, gas utilization, permeability-related indices, and spatial thermal states. Particular attention is given to the integration of metallurgical mechanisms with data-driven methods, including process-informed features, mechanistic surrogates, residual learning, physical constraints, temporal models, multimodal fusion, domain adaptation, and uncertainty quantification. The analysis shows that reported prediction errors cannot be meaningfully compared without consistent definitions of the target variable, forecast horizon, information available at prediction time, and validation protocol. It further distinguishes temporal robustness within a single furnace from cross-condition and cross-furnace generalization, emphasizing the risks of information leakage, negative transfer, interpolated labels, and delayed assay feedback. Publicly available industrial datasets provide a feasible basis for reproducible silicon-content forecasting, but their use remains constrained by incomplete assay provenance, uncertain result-release times, sensor validity, and limited cross-furnace coverage. A release-aware benchmarking framework is therefore proposed, incorporating chronological evaluation, matched baselines, mechanism-specific ablation studies, and calibrated prediction intervals. Future progress depends not merely on increasingly complex model architectures, but on demonstrating which forms of metallurgical knowledge provide reliable predictive value under clearly defined operating shifts and real-time information constraints.