Predicting Temperature of Carbon Brick in a Blast Furnace Using Machine Learning Approaches

The blast furnace is crucial for iron production in industries, with its campaign life directly impacting iron-producing costs. One significant factor affecting its lifespan is the thickness of the carbon brick which is closely related to the hearth temperature. However, the intensive smelting leads to abnormally high temperatures that erode the carbon bricks and shorten the blast furnace's lifespan. We proposed a machine-learning (ML) model for predicting the hearth temperature to solve this issue. In this study, a Long Short-Term Memory (LSTM) model was trained to predict the hearth temperature. The training data set originated from the operation data of one of the blast furnaces for 2761 days each of which contained 1470 features. After discussions with CSC experts, the feature number was reduced from 1470 to 360 by removing irrelevant features. Next, the top 150, 100, and 50 features related to the hearth temperature were found using the Pearson correlation coefficient. Three LSTM models were trained using three feature subsets to optimize the feature number. Furthermore, different combinations of input and output lengths were tested to optimize the model. The input lengths were 15, 30, and 45 days, and the output lengths were 1, 3, and 5 days. The last 480 days were separated from the training dataset to examine the long-term prediction of the proposed LSTM model. Since the working conditions, raw materials, and operation protocols were mutated during operation, the frequency of updating the prediction model was investigated to improve the long-term prediction accuracy. The dataset with 150 features achieved the best performance with a mean squared error (MSE) of 0.0829. For the optimal configuration, the previous 30 days' features were used to predict the temperature for the next 3 days, updated every three days. This configuration achieved the lowest MSE of 0.00939, much better than the average MSE of all groups of 0.0214. The best combination of the dataset and machine learning (ML) model was selected as a result of this study.