Hot Metal Temperature Prediction Technique Based on Feature Fusion and GSO-DF

Abstract

Hot metal temperature was a direct indicator of blast furnace condition. If the operator predicted its trend in advance, it was conducive to the stable operation of the blast furnace. This study combined expert experience and big data technology to propose an intelligent prediction method for hot metal temperature. Based on metallurgical theory and data governance algorithms, outlier processing, data simplification, data standardization and frequency unification of blast furnace data were completed. The blast furnace feature was processed by multiple feature engineering method, 8 rammed residual blast furnace features were eliminated; 36 features were screened by feature selection technique to form the optimal combination of hot metal temperature prediction; 4 derived parameters were constructed by PCA technique. Applying the filtered combination of feature as input, a GSO-DF model was created, which was satisfactory in predicting the hot metal temperature in the next hour. The MAE and MSE of the GSO-DF model was 3.54 and 27.34, respectively. It achieved a hit rate of 92.86% in the ±10°C range. The average hit-rate of the model can reach more than 91% by updating the model every day to test the data of the coming month. Even if the hot metal temperature fluctuated greatly, it was still able to predict the temperature trend well and provide reliable guidance for the field personnel. The hot metal temperature qualified rate increased by 6.8% during the model application period. It contributed to the improvement of hot metal quality at the site, and achieved satisfactory result.