(765b) High Performance Prediction of Molten Steel Temperature Through Gray-Box Model

The steel industry faces a stiff competition in the global market; each steel company has to produce high quality products satisfying various customers demand. A key problem is variation in product quality. For example, surface cracks may occur and product quality may deteriorate if molten steel temperature is not controlled precisely in a continuous casting process. In the steel making process, molten steel temperature in the tundish is one of the key factors causing variation of product quality. It is crucial to develop a prediction and control system of molten steel temperature to achieve high quality steel production. However, the molten steel temperature cannot be directly controlled in the continuous casting process because there is no effective manipulated variable. In order to realize precise temperature control, it is necessary to adjust the molten steel temperature at the end of secondary refining. Therefore, a model relating the molten steel temperature in the continuous casting process and in the secondary refining process needs to be constructed. The model development is difficult because various factors, e.g. degradation of ladles and treating time at each station, affect molten steel temperature in the ladle and the tundish.

In the present research, a novel gray-box model is proposed to predict the molten steel temperature by combining a first-principles model and a statistical model. First, the first-principles model was developed for a converter, a secondary refining process, and a continuous casting process. This model takes account of heat balance among molten steel, slag, air, ladle, and tundish; the results of computational fluid dynamics (CFD) analysis were used to simplify the model and to achieve high prediction performance. Then, the statistical model was developed to compensate the prediction error of the first-principles model, because the prediction performance of the first-principles model was not sufficient for its industrial application. Finally, the temperature prediction is made by adding the output of the first-principles model and that of the statistical model.

To validate the developed prediction system, various types of models were compared: the first-principles model, a partial least squares (PLS) model as a linear model, locally weighted PLS (LW-PLS) model as a just-in-time model, a random forest (RF) model as a nonlinear model, and the gray-box models combining them. The prediction performance was evaluated by using real operation data provided by Sumitomo Metal Industries, Ltd. in Japan. As a result, it was confirmed that the gray-box model combining the first-principles model and RF could achieve the best prediction accuracy. Its root mean square error (RMSE) was improved by 35% and 37% in comparison with RF and the first-principles model, respectively. PLS and LW-PLS cannot realize the better prediction than RF. These results have clearly shown the advantage of the developed gray-box model over the other models. The gray-box model is now used to investigate the advanced control system of molten steel temperature.