基于机理约束神经网络的LF精炼终点温度预测

he LF refining process, widely used as the final step in steelmaking, highlights the critical importance of endpoint temperature control. Presently, most steel mills rely on operators\' experience and multiple temperature measurements to regulate endpoint temperatures, encountering challenges in achieving precise prediction and control. This paper focuses on a 260t LF furnace in a domestic steel mill, initiating with a thermal balance analysis of the refining process to identify factors influencing the endpoint temperature. It is discovered that these factors directly impact the heating range (the difference between outlet and inlet temperatures during LF refining), thereby indirectly affecting the endpoint temperature. Consequently, in modeling, the heating range is adopted as the output for indirect endpoint temperature prediction. By combining metallurgical principles with data analysis, two predominant factors affecting the LF heating range are uncovered: the initial LF temperature and the specific electric energy consumption per ton of steel, along with their monotonic relationships with the heating range. Guided by the principle of monotonic constraint, these relationships are incorporated into the loss function of a neural network, establishing a single-hidden-layer neural network model. This model achieves an accuracy rate of 94.05% within a ±10 C error margin, surpassing the 91.00% accuracy when using artificial neural networks for direct endpoint temperature prediction and the 92.14% accuracy in indirect endpoint temperature prediction via artificial neural networks.