Development of a GCN-LSTM based prediction model for sticking breakout in continuous casting

Continuous casting is a critical stage in steelmaking; however, it is prone to breakout accidents (especially sticking breakouts) owing to complex influencing factors and mechanisms. Therefore, it is necessary to develop an effective prediction model to help improve operational safety. This study conducted an in-depth analysis of the sticking breakout mechanism, tracing the complete process from sticker initiation to crack propagation, and clarified the spatial characteristics of the thermocouple temperature variations during such events. A novel deep-learning framework was developed, which utilizes the first-order derivative of the thermocouple temperature as the model input. This framework integrates long short-term memory (LSTM) network and graph convolutional network (GCN) models to learn and extract spatiotemporal features from temperature data. Through designed ablation experiments and an extensive analysis of field data to investigate the impact of the model parameters, the optimal parameters were identified, and the effectiveness and superiority of the GCN-LSTM model for sticking breakout prediction were verified. The final optimal model, with the number of iterations, learning rate, time step, and the number of hidden layer dimensions set to 176, 0.01, 20, and 5, respectively, achieved a prediction rate, correct alarm rate, and F₁-score of 97.93%, 98.75%, and 98.34%, respectively.