Prediction of carbon content and temperature at the end of converter steelmaking based on Just-in-time learning

Accurate carbon content and temperature prediction at the endpoint of converter steelmaking can provide essential reference values for the control of the endpoint. In order to achieve accurate prediction, a kind of mutual information (MI) weighted Just-in-time learning (JITL) based on Kmeans clustering was proposed, aiming at the characteristics of high dimensionality, nonlinearity, and high volatility of converter steelmaking process data. The JITL modeling method predicts the carbon content and temperature at the end of converter steelmaking. Firstly, the improved Pelican optimization algorithm (MPOA) was used for feature selection, reduced data dimension, and critical factors affecting carbon content and temperature at the endpoint were screened out as input to the model. Secondly, the input sample data was clustered to reduce the impact of data fluctuations on the model. Finally, by weighting the mutual information of the prediction and training samples, the local algorithm learning set is constructed according to the degree of contribution of the samples. It is used for the local model training of the hybrid kernel Extreme Learning Machine (HKELM) hyperparameters optimized by MPOA. Finally, the carbon content and temperature are predicted. The experimental results show that the hit rate of carbon content is 95% at ±0.03%, and the hit rate of temperature is 97% at ±15 °C.