Regular Article

Performance analysis of hot metal temperature prediction in a blast furnace and expert suggestion system proposal using neural, statistical and fuzzy models

¹ Karabük University, Şenpiliç, Barbaros Mah. Kardelen Sok. No: 2/68, Palladium Tower, Ataşehir, İstanbul, Turkey
² Department of Computer Engineering, Faculty of Engineering, Karabük University, Balıklarkayası Mevkii, Karabük 78050, Turkey
³ Honeywell, Çayır Cad. Üçgen Plaza No: 7 Kat: 5 Ataşehir, İstanbul 34752, Turkey

^* e-mail: yasin.tunckaya@honeywell.com

Received: 6 February 2021
Accepted: 3 May 2021

Abstract

Blast Furnace (BF) production methodology is one of the most complex process of iron & steel plants as it is dependent on multi-variable process inputs and disturbances to be modelled properly. Due to expensive investment costs, it is critical to operate a BF by reducing operational expenses, increasing the performance of raw material and fuel consumptions to optimize overall furnace efficiency and stability, also to maximize the lifetime. The chemical compositions and temperature of hot metal are important indicators while evaluating the operation, therefore, if the future values of hot metal temperature can be predicted in advance instead of subsequent measuring, then the BF staff can take earlier counteractions on several operational parameters such as coke to ore ratio, distribution matrix, oxygen enrichment rate, blast moisture rate, permeability, flame temperature, cold blast temperature, cold blast flow and pulverized coal injection rate, etc. to control the furnace optimally. In this study, Artificial Neural Networks (ANN) model is proposed combined with NARX (Nonlinear autoregressive exogenous model) time series approach to track and predict furnace hot metal temperature by selecting the most suitable process inputs and past values of hot metal temperatures using the real data which is collected from the BF operated in Turkey during 2 months of operation. Various data mining techniques are applied due to requirements of charge cycling and operating speed of the furnace which secures novelty and effectiveness of this study comparing previous articles. Furthermore, a statistical tool, Autoregressive Integrated Moving Average (ARIMA) model, is also executed for comparison. ANN prediction results of 0.92, 8.59 and 0.41 are found very satisfactory comparing ARIMA (1,1,1) model outputs of 0.73, 97.4 and 9.32 for R² (Coefficient of determination), RMSE (Root mean squared error) and MAPE (Mean absolute percentage error) respectively. Consequently, an expert suggestion system is proposed using fuzzy if-then rules with 5 × 5 probability matrix design using the last predicted HMT value and the average of the last 5 HMT values to decide furnace’s warming or cooling movements state in mid-term and maintain the operational actions interactively in advance.