Sintering yield prediction model based on time series feature extraction and composite neural network

Abstract

Sintering is an important process in the iron and steel industry. As the main raw material of blast furnace, the proportion of sinter is more than 70%. It is also an important work for steel plants to ensure and improve sintering yield. Sintering yield is inherently prone to significant variations, primarily driven by the quality of raw materials, the condition of equipment, and the intricacies of production operations. Additionally, the considerable time lag between changes in the ratios of raw materials, operational parameters and the ultimate sintering yield leaves a gap in the operational decision-making process. To address this, the present study employs time series correlation analysis to extract both dynamic and static characteristics intrinsic to sintering yield. In view of the advantages of CNN–LSTM (convolutional neural network–long short-term memory) in processing time-series data, we leverage a CNN–LSTM composite algorithm and a BP neural network algorithm to extract feature information in dynamic and static characteristics. The resulting sintering yield prediction model was rigorously trained and tested using actual sintering yield data from a large-scale iron and steel production enterprise. This model achieved a determination coefficient (R²) of 0.96, showcasing an impressive accuracy of 86.11% within an absolute error range of ±10 t·h⁻¹ in forecasting the yield of finished products across a 72-hour continuous production cycle. The resluts indicate that the proposed approach significantly enhances sintering yield forecasting accuracy, providing a valuable tool for optimising production planning and decision making in the steel industry.