Real-Time Carbon Content Prediction Model for the Reblowing Stage of Converter Based on PI-LSTM

TL;DR

This paper introduces a new model for predicting carbon content in steelmaking that combines physics and data to improve accuracy and real-time control.

Contribution

The novel PI-LSTM model integrates physical constraints with data-driven learning for real-time carbon prediction in steel converters.

Findings

The PI-LSTM model achieved an MAE of 0.0077 and RMSE of 0.0112, outperforming traditional models.

Endpoint carbon hit rates within ±0.005% reached 53.71%, showing improved prediction accuracy.

The model enhances physical plausibility and supports real-time control in steel production.

Abstract

Precise forecasting of carbon content in the converter’s reblowing phase is pivotal to boosting steel production efficiency and ensuring effective control over molten steel quality. However, existing mechanistic models based on material balance and decarbonization kinetics suffer from insufficient accuracy due to simplifying assumptions. In contrast, data-driven models rely on data quality, lack generalization capability, and lack physical interpretability. Additionally, integral models based on flue gas analysis suffer from data latency issues. To overcome these limitations, this study proposed a real-time carbon content prediction model for the converter’s reblowing phase, leveraging a physics-informed long short-term memory (PI-LSTM) network. First, flue gas data was processed using a carbon integration model to generate a carbon content change curve during the reblowing stage as a reference for actual values; second, a dual-branch network structure was designed, where the LSTM branch simultaneously predicts carbon content and key unmeasurable parameters in the decarbonization kinetics, while the mechanism branch combined these parameters with the decarbonization formula to calculate carbon content under mechanism constraints; finally, a joint loss function (combining data-driven loss and mechanism constraint loss) was used to train the model, and the gray wolf optimization (GWO) algorithm was employed to optimize hyperparameters. Experimental results show that compared to the mechanism model (MM) and LSTM model, the PI-LSTM model achieves an average absolute error (MAE) of 0.0077, a root mean square error (RMSE) of 0.0112, and endpoint carbon content hit rates within ±0.005%, ±0.01%, ±0.015% error ranges, achieving 53.71%, 82.23%, and 95.45%, respectively, significantly improving prediction accuracy and physical plausibility. This model lays a robust groundwork for dynamic closed-loop real-time control of carbon levels in the converter’s reblowing stage.

Figures