Method to predict alloy yield based on multiple raw material conditions and a PSO-LSTM network

Abstract

The production of ferroalloys accounts for a large proportion of the total energy consumption of the steelmaking industry. Accurately predicting alloy element yields is the key to reducing alloy waste, but there are significant differences in alloy yield under different conditions using ferroalloy raw materials during steelmaking. Therefore, this paper proposes a multi-model alloy element yield prediction method based on a particle swarm optimization (PSO) hyperparameter-optimized long short-term memory (LSTM) network and raw material condition classification. The accuracy of the PSO-LSTM prediction model was verified through simulations and was significantly higher than that of other network models when using the same raw material conditions. The average absolute error of predictions using raw materials with a low drum index was 0.4485, and it was 0.6162 under a high drum index, which was significantly lower than that (0.7077) under the condition without classification. This demonstrates the rationality of classifying working conditions according to the raw material conditions of ferroalloys. In addition, this paper combines the prediction model with a linear programming algorithm to develop a ferroalloy operating system and uses it in a steel plant to guide workers to complete an alloying operation. After four months of industrial testing, the internal control rate of finished steel composition increased from 91–94 % to 95–98 %. According to statistical analysis, the optimized HRB400E threaded steel consumed 1.23 kg less silicon-manganese per ton of steel, reduced the cost of steel alloy per ton by 8.6 yuan, and significantly reduced the waste of ferroalloys during steelmaking.

Keywords

• Steel industry
• Ferroalloy
• LSTM
• PSO
• Machine learning