A New Incremental Learning Modeling Method Based on Multiple Models for Temperature Prediction of Molten Steel in LF

作者：Tian Huixin^*; Mao Zhizhong; Wang Anna

来源：ISIJ International, 2009, 49(1): 58-63.

DOI：10.2355/isijinternational.49.58

• ladle furnace
• incremental learning
• multiple models
• temperature prediction

Traditional temperature prediction models of molten steel have to face the dilemma problems about updating models. Aiming at these problems, a new incremental learning modeling method based on multiple models is proposed in this paper. Firstly, an intelligent model based on ELM is established by analyzing the conversation of energy during whole refining process of LF as generic sub intelligent model. Secondly, the errors of different generic sub intelligent models are calculated. The weights of sub intelligent models can be obtained by these errors. Then the temperature prediction model is presented by aggregating these sub intelligent models. Finally, when new production data accumulate enough, they will be used to train a new sub intelligent model and the sub model's weight will be obtained according to the errors of training. Then the new aggregated temperature prediction model is established based on all of the generic sub intelligent models that include the old ones and the new one. Till then, the updating of prediction model is completed. The new incremental learning method preserves the information of old sub models by this process, and no longer by saving the all original data, Therefore, it is efficient to save a mass of space and time. The new temperature prediction model with incremental learning is used in 300 t LF of Baoshan Iron & Steel Co. Ltd. The results demonstrate both updating ability and accuracy of new model are satisfied for production.

• 出版日期2009
• 单位东北大学