Prediction of blast furnace hot metal temperature based on support vector regression and extreme learning machine

The hot metal temperature is a key process parameter for blast furnace (BF) ironmaking that reflects the quality of hot metal, the thermal state of BF hearth, the energy utilization efficiency of BF, and many other information. Prediction of the hot metal temperature in the next smelting cycle will be helpful in gaining a better understanding of the change trend of hot metal quality and BF smelting status in time. With this, corresponding operational measures can be conducted to maintain the BF stable and smooth state, high production, and low consumption. Nowadays, big data technology has made considerable progress toward a more accurate and faster collection, storage, transmission, query, analysis, and integration of mass data, providing a good data foundation for data-driven machine learning models. In addition, with the substantial increase in computer calculation speed and the significant development of algorithms, the prediction accuracy of deep machine learning models has noticeably improved. The development of these technologies provides feasibility for the prediction of important indicators under complex industrial conditions. Based on the data produced from a 4000-m³ BF in a large span time range (2014–2019) and daily time dimension, this paper considered hot metal temperature as the objective function. First, the characteristic parameters of hot metal temperature were processed by linear and nonlinear correlation analysis, feature selection, and normalization methods. Then, the positive and negative correlation characteristic parameters that have a significant influence on the temperature of the hot metal were obtained. Finally, prediction models of hot metal temperature were established based on two algorithms of support vector regression and extreme learning machine. Although both the algorithms can achieve effective prediction, results from support vector regression are better at an average absolute error of 4.33 °C and a hit rate of 94.0% (±10 °C).