A machine learning approach for intra- and inter-cast hot metal temperature forecasting

Hot metal temperature (HMT) forecasting is beneficial for safe and efficient operation of blast furnaces (BFs) in the iron and steel industry. Despite relevant literature, most works addressing HMT forecasting focus on specific solutions tailored for a given (private) dataset. Such datasets reflect particular features of a blast furnace, posing challenges in determining the optimal strategies for HMT forecasting. To tackle the scarcity of more comprehensive studies, this paper presents a more in-depth quantitative and qualitative analysis of machine learning models for HMT forecasting. The study involved over 80,000 temperature records and focused on evaluating the performance of models trained specifically for intra-cast (short-term) and inter-cast (longer-term) scenarios. A detailed temporal analysis revealed that the Multilayer Perceptron yielded the lowest Mean Absolute Error for two-cast-ahead prediction: 11.280 °C for a prediction horizon between 5.5 to 6 h. Considering a variety of horizons, however, simpler and more time-efficient methods, such as Linear Regression and Partial Least Squares, yielded better performance in most cases. The findings offer valuable insights for better decision-making in BF control and maintenance.