LF Molten Steel Temperature Control With Interpretable ML

The patent introduces SHAP (SHapley Additive exPlanations) as an intermediary layer between the hybrid prediction model and the operator. SHAP values decompose the complex model output into interpretable contributions from each input feature, allowing operators to understand which parameters (e.g., heating time, alloy composition) most influence temperature predictions without simplifying the underlying complex model architecture.

The patent replaces traditional opaque mathematical modeling approaches with a data-driven machine learning model enhanced by game-theory-based SHAP explanations. This substitution allows the system to capture complex nonlinear relationships in metallurgical processes while providing interpretable results through SHAP's cooperative game theory framework, which attributes prediction outcomes to specific input features in a mathematically rigorous manner.

The patent implements a closed-loop feedback control system where the hybrid prediction model continuously forecasts molten steel temperature based on real-time process parameters (heating time, power input, alloy composition). These predictions feed back to operators who adjust control parameters accordingly, creating a continuous cycle of prediction-adjustment-verification that stabilizes temperature control while maintaining operational flexibility.

The patent applies preliminary action by using the prediction model to forecast temperature outcomes before actual processing occurs. Operators receive advance predictions of temperature evolution based on planned parameter settings, allowing them to pre-adjust control parameters to achieve target temperatures more accurately and reduce fluctuations during the actual heating process.

The patent merges two complementary approaches: (1) a metallurgical mechanism model that incorporates physical and chemical process knowledge (providing interpretability and feasibility), and (2) a machine learning model that captures complex nonlinear relationships from data (providing accuracy). The ensemble hybrid model combines predictions from both components, leveraging the strengths of each to achieve both accuracy and construction feasibility.

The patent creates a composite modeling approach analogous to composite materials, where the hybrid model integrates heterogeneous components (mechanism-based predictions and data-driven predictions) with different properties. Each component contributes unique strengths: the mechanism model provides physical consistency and interpretability, while the machine learning component provides adaptive accuracy. Their combination creates a superior integrated model that overcomes the limitations of individual approaches.