A real-time temperature prediction method based on CNN-LSTM with MODE in steelmaking process

Prediction of molten steel quality is a key issue in the converter steelmaking process. Due to high temperature and physical-chemical reaction, it is difficult to predict the temperature of molten steel in real time through mechanism analysis. This is a challenging problem with black box characteristics. Here we propose a Convolutional neural networks (CNN) and long short-term memory (LSTM) networks model based on multi-objective differential evolution algorithm (MODE) for the real-time prediction problem in the steelmaking process. Specifically, the CNN is used to extract feature between multiple variables that affect the temperature of molten steel, and the LSTM is suitable for the time series model based on the irregular trends information. The multi-step output CNN-LSTM method is used to construct the temperature prediction model of molten steel, and a knee solution strategy is applied to MODE , which can balance the model accuracy and generalization ability. The experiments were conducted using real data collected from the converter steelmaking process. The results verify the effectiveness of the proposed model on multiple steel grades and outperform other artificial intelligence methods in multi-task prediction.