Multi-time-scale TFe prediction for iron ore sintering process with complex time delay. (August 2019)

• Record Type: Journal Article
• Authors: Chen, Xiaoxia; Shi, Xuhua; Tong, Chudong
• Is Part Of: Control engineering practice. Volume 89(2019)
• Page Start: 84
• Page End: 93
• Publication Date: 2019-08
• Subjects: Data-driven model -- Iron grade -- Multi-scale prediction -- Time delay -- Iron ore sintering; Automatic control -- Periodicals; 629.89
• DOI: 10.1016/j.conengprac.2019.05.012
• Languages: English
• ISSNs: 0967-0661

Abstract

Abstract: Iron ore sintering is a critical process in steel-making industry. It produces sinter with qualified iron grade (TFe) for the blast furnace process. The process variables display a multi-time-scale feature that is derived from the complex time delays involved in a sintering process. To resolve the contradiction between a single-time-scale data acquisition of TFe and a multi-time-scale feature that the process displays. A multi-scale prediction approach in offline and online modes was presented for the prediction of TFe. The approach not only solves the contradiction, but also meets the requirements of online and offline optimizations of a sintering process. First, a discrete wavelet transform method was combined with process knowledge to decompose the online and offline TFe component with different time scales. Then, an improved back-propagation neural network (IBPNN) with its input neurons not only connected to the hidden neurons but also to the output neurons was built for the offline TFe prediction under large-time scale. Last, a just-in-time-learning-based online model was built under the mixture time scales of medium and small. The simulation results of actual run data show that an IBPNN has a good overall performance compared with an extreme learning machine (ELM), an improved ELM, and a back-propagation neural network for the offline TFe prediction. The results also show the superiority of the multi-time-scale prediction model compared with an offline prediction and a one-time-scale prediction models.