Abstract: Combining the conventional mechanism method with the newly developed intelligent algorithm, a hybrid model was developed for predicting the molten steel temperature in LF. The mechanism method is used to build the thermal model by analyzing the energy going into and out of the molten steel during the LF refining production process. It is hard to calculate the coefficients in the thermal model by mechanism method, so they are estimated by experience traditionally. In this paper, a new ensemble ELM algorithm using the modified AdaBoost.RT is proposed to calculate these coefficients. The new hybrid model overcomes the difficulty of obtaining the coefficients in thermal model, and solves the problem of limited prediction precision using “black box” model. The experiment demonstrates that the new hybrid model can improve generalization performance and the prediction accuracy.
Introduction
In modern steelmaking, the Ladle Furnace (LF) plays a critical role in refining molten steel to the precise temperature and chemical composition required for continuous casting. Accurate temperature prediction is essential; deviations can lead to significant production delays, energy waste, or compromised product quality. However, modeling the LF environment is notoriously difficult due to the extreme temperatures (reaching 1,600°C) and the complex interplay of physical and chemical reactions.
Engineers typically face a "Goldilocks" problem when choosing a modeling approach:
- Mechanism Models: These are based on the laws of thermodynamics and energy conservation. While physically sound, they rely on specific coefficients—such as the thermal efficiency of arcing or heat loss through refractory walls—that are nearly impossible to measure directly in a production environment.
- Black Box Models: Pure machine learning models, such as standard neural networks, can learn patterns from data but often ignore underlying physics. This lack of structural guidance can lead to poor generalization when furnace conditions shift outside the training data range.
The Hybrid Modeling Approach
To address these limitations, this paper proposes a hybrid model that leverages the strengths of both physical laws and data-driven intelligence. Instead of using machine learning to predict the final temperature directly, the "intelligent algorithm" is used to estimate the elusive physical coefficients required by the thermodynamic equations.
The core of this architecture consists of two integrated parts:
- Thermal Mechanism Model: A physical framework based on energy conservation accounting for heat gain from arcing, heat loss through refractory walls, the cooling effect of alloy/slag additions, and surface heat loss.
- Ensemble Extreme Learning Machine (ELM): The model utilizes an ensemble of ELMs to calculate coefficients like arcing efficiency ($\eta$) and surface heat loss ($f_{Surf}$). ELMs are chosen for their high training speed and ability to avoid the "local minima" problems common in traditional Back-Propagation Neural Networks (BPNN).
To enhance the stability of the ELM, the researchers employed a modified AdaBoost.RT algorithm. This includes a self-adaptive threshold that automatically adjusts based on the root mean square error of each iteration, preventing the model from overfitting to noise while ensuring it learns from challenging data points.
Experimental Results and Discussion
The model was validated using 250 production data points from a 300-ton Ladle Furnace at Baoshan Iron & Steel Co. Ltd. The dataset included variables such as refining power consumption, initial temperature, ladle states, and the weight of molten steel.
The hybrid model demonstrated significant improvements in both precision and generalization. Specifically, it achieved a predictive error of $\le$ 5°C in over 90% of the test cases, meeting the rigorous industry standard for successful metallurgical control.
As shown in the error distribution analysis, the ensemble ELM hybrid model's errors are more tightly clustered around zero compared to both the single ELM "black box" and the hybrid model using a BPNN. This confirms that the ensemble approach effectively stabilizes the predictions and that the hybrid structure provides better grounding than a pure data-driven model.
Conclusion
By letting machine learning handle the "messy," unmeasurable variables within a structured physical framework, the hybrid model provides a robust solution for industrial temperature prediction. While the current model was tuned for a specific 300t furnace, the methodology is universal. Future work may extend this approach to other complex, high-temperature industrial processes where physical parameters are difficult to measure, such as blast furnaces or basic oxygen furnaces.
Summary updated