基于IPSO-RELM转炉冶炼终点锰含量预测模型

分析了影响转炉冶炼终点钢水中锰含量的因素, 针对基于BP神经网络算法的转炉冶炼终点锰含量预测模型存在的收敛速度慢, 预测精度低等问题, 提出了一种基于极限学习机(ELM) 算法建模的新思路, 并引入正则化以及改进粒子群优化算法(IPSO), 建立了基于改进粒子群算法优化的正则化极限学习机(IPSO-RELM) 的转炉终点锰含量预测模型; 应用国内某炼钢厂转炉实际生产数据对模型进行训练和验证, 并与基于BP、ELM和RELM算法的三类模型进行比较.结果表明, 采用IPSO-RELM方法构建的模型, 锰含量预测误差在±0. 025%范围内的命中率达到94%, 均方误差为2. 18×10-8, 拟合优度R2为0. 72, 上述三项指标均显著优于其他三类模型, 此外, 该模型还具有良好的泛化能力, 对于转炉实际冶炼过程具有一定的指导意义. 

Improved prediction model for BOF end-point manganese content based on IPSO-RELM method

The basic oxygen furnace (BOF) steelmaking process, as the predominant steelmaking method used around the world, involves very complex physical and chemical phenomena such as multi-component reactions, multi-phase fluid dynamics, and high temperature. The main task of the BOF process is tailoring the temperature and melt components to meet the requirements of high-quality steel production. With the development of intelligent steelmaking, the prediction of the end-point manganese content is an extremely important task for the BOF process, and improving the level of control regarding the end-point of BOF steelmaking can reduce production costs and enhance efficiency. In this paper, the mechanism of the BOF steelmaking process and the factors influencing the endpoint manganese content were analyzed. The control variables for predicting the end-point manganese content were also determined. To solve the problems of slow convergence, weak generalization ability, and low prediction accuracy in the prediction model established for the BP neural network, a new modeling concept based on an extreme learning machine (ELM) algorithm was proposed. By introducing regularization and improved particle swarm optimization (IPSO), a prediction model for the end-point manganese content in a converter based on improved particle swarm optimization and a regularized ELM (IPSO-RELM) was established. The paper then trained and verified the performance of these models with actual production data. A comparison of the performance of the proposed model with those of the prediction model of the BP neural network, the ELM model, and the RELM model reveals that the IPSO-RELM prediction model has the highest prediction accuracy and the best generalization performance. The hit ratio of the IPSO-RELM prediction model is 94%when the predictive errors of the model are within 0. 025%, the mean square error is 2. 18 × 10-8, and the fitting degree is 0. 72. Relative to the above three models, the IPSO-RELM prediction model may provide a more accurate prediction of the end-point manganese content and thus serves as a good reference point for actual production.