Abstract: Continuous real-time prediction of carbon and temperature in converter steelmaking process is the key to the end-point control.In view of the large difference of time series samples between furnaces and affecting the accuracy of the model prediction,a new method for predicting the end-point carbon temperature by using real-time learning strategy of quadratic similarity measurement with a Long-Short Term Memory (LSTM) network hybrid model that had feedback compensation mechanism were presented.The initial local model was established by using grey relational degree criteria,and the centroid sample was located in the local model to obtain the optimal local model.The output feedback of LSTM network was backed to the input in training,which could improve its generalization ability in the secondary training.The experimental results showed that the accuracy of temperature prediction error was 76.7% and carbon prediction error was 86.7% according to the process requirements of steelmaking.

Key words: converter steelmaking, instant learning, grey relational degree, long short-term memory network

摘要： 转炉炼钢过程中碳、温连续实时预报是终点控制的关键,针对炉次间的时间序列样本差异度大进而影响模型预测精度的问题,提出一种二次相似性度量的即时学习策略加具有反馈补偿机制的长短期记忆网络混合模型预测终点碳温方法。利用灰色关联度准则进行一次度量建立初始局部模型,在其局部模型中定位出“质心”样本进行二次度量,从而获取最优局部模型;长短期记忆网络训练时,将输出反馈回输入端,使其二次训练能够提高泛化能力。通过钢厂转炉炼钢生产过程的实际数据对方法进行测试和验证,结果表明,按照炼钢工艺要求,温度预测误差在±5℃的精度为76.7%,碳预测误差在±0.01的精度为86.7%。

关键词: 转炉炼钢, 即时学习, 灰色关联度, 长短期记忆网络