Dynamic, data-driven prediction model of molten steel temperature in the second blowing stage of a converter

Molten steel temperature is a parameter in converter end-point control. Accurate prediction of molten steel temperature is crucial for converter end-point control. However, most of the previous end-point prediction models are static models, which can only predict the molten steel temperature at the end-point of converter blowing and cannot realize dynamic prediction, affording a limited role for these models. To solve this challenge, a data-driven prediction model of molten steel temperature in the second blowing stage in a converter is proposed. First, the model retrieves the similar cases in the historical case base through the process parameters in the main blowing stage of the new case, such as carbon content and temperature of TSC measurement, based on the case-based reasoning (CBR) algorithm. Second, the process parameters in the second blowing stage of the similar cases, such as oxygen flow, lance position, and argon flow, are used to train the relationship between the process parameters and the molten steel temperature based on the long short-term memory (LSTM) algorithm. Third, the trained LSTM model is used to dynamically calculate the molten steel temperature in the second blowing stage of the new case. Finally, the actual production data is divided into five sets for cross-validation, and the model prediction accuracy changes are tested when the number of reuse cases ranges from 1 to 10, and the number of neurons is 5, 10, 15, and 20. The results show that, on the one hand, the prediction accuracy of the model first increases and then decreases with an increasing number of cases, and when the number of reused cases is 4, the prediction accuracy of the model is the highest, indicating that the number of cases is increased when training the model. Improving the prediction accuracy of the model is beneficial; however, the reference value of the case decreases with the similarity of the case, reducing the prediction accuracy of the model. Conversely, when the number of neurons is 10, the prediction accuracy of the model reaches it’s the highest value. The hit rate of the prediction error in the range of [−5 ℃, 5 ℃], [−10 ℃, 10 ℃], and [−15 ℃, 15 ℃] reached 40.33%, 68.92%, and 88.33%, respectively. This paper also establishes the traditional quadratic model and cubic model as well as further proves the effectiveness of the model by comparing the three indicators of these models, namely, the RMSE, MSE, and hit rate.