A multi-task learning prediction model for the endpoint of converter steelmaking based on genetic algorithm optimised

The temperature and composition of molten steel are important indicators that affect its quality, and a high-precision endpoint prediction model is needed for intelligent smelting using converters. This study addresses the current neglect of the correlation between objectives (temperature and composition) and the need to improve the accuracy of objective prediction. Thus, the potential relationship between molten steel temperature and carbon content is analysed through the metallurgical mechanism and energy balance. A multi-task neural network learning model based on a genetic algorithm, which is used to optimise the weights of loss functions for each task in a multi-task prediction model, is proposed for the high-accuracy simultaneous prediction of the endpoint carbon content and temperature of molten steel in a converter. The model was validated using actual production data from a steel plant. The results showed that within carbon content and temperature error ranges of [−0.02%, 0.02%] and [−15°C, 15°C], respectively, the carbon content–temperature double-hit rate of the model before and after using the genetic algorithm optimised increased by a maximum of 22.5% (W (0.71,0.29) and W (0.6,0.4) ). Compared with the hit rates of single-objective prediction models (genetic algorithm–back propagation neural network, case-based reasoning, and multiple linear regression), within the same error ranges, the hit rate of the proposed model is better by 6.2%, 17.5%, and 7.5%, respectively. This indicates that the developed model simultaneously predicts the endpoint carbon content and temperature of molten steel in a converter with high accuracy, providing a reference for the accurate prediction of converter endpoints.