Data-based residence time estimation for continuous process data series based on the example of a sinter plant for the prediction of return fines

RWTH Publications (RWTH Aachen)

Abstract

The Blast Furnace - Basic Oxygen Furnace (BF-BOF) route remains the dominant method of crude steel production, accounting for 71 % of global crude steel production in 2022. Within this process, iron ore sinter plays a crucial role as a primary feed material. Since iron ore fines must be agglomerated into particles of 5–50 mm to ensure sufficient gas permeability in the blast furnace, sinter plants are essential for agglomeration but also for recycling process residues. Producing high-quality sinter is therefore both economically and environmentally important. However, sinter quality is affected by multiple factors throughout production, including mixing, ignition, sintering, crushing, and cooling. Additionally, transport and storage degrade the material, generating fines and segregation. A major challenge is that sinter cannot be tracked through the process, and quality parameters cannot be measured in real time. Instead, manual testing introduces delays of several hours, forcing operators to rely on experience. This can lead to inefficient operation, higher emissions, and inconsistent product quality. Although modern control systems like PLCs and DCS collect large amounts of process data (e.g., temperature, pressure, mass flow), these data are time-based and not linked to specific production units. A key difficulty in modelling sinter quality is the varying Residence Time (RT) of particles as they move through the plant. These Residence Time Distributions (RTD) change dynamically due to process conditions and material behaviour such as mixing and segregation, especially in components like the sinter shaft cooler. Traditional RTD estimation methods rely on tracer experiments, which are difficult to apply in sinter plants due to extreme temperatures. Alternative approaches, including neural networks or manually labelled datasets, either lack interpretability or fail to capture short-term dynamics effectively. To address this, the thesis introduces a novel data-driven Residence Time Estimation (RTE) method. This approach aligns time-series process data with the actual movement of sinter material, enabling effective application of Machine Learning (ML) models. By iteratively testing different RT assumptions and evaluating prediction accuracy, the method identifies the best matching residence time using an ML-based correlation function. Applied to real plant data, the method successfully estimated RTDs and improved prediction of sinter quality indicators such as Cold Return Fines (CRF). Using aligned data, models achieved prediction accuracies up to R² = 0.60 and enabled forecasts up to 73 minutes in advance. This allows earlier process adjustments, improving efficiency, reducing sinter fines, and enhancing blast furnace performance. Even small improvements have significant economic and environmental impact: reducing return fines can lead to substantial cost savings while simultaneously lowering energy consumption and CO₂ emissions. Overall, the proposed method represents a major advancement in process analysis by enabling material tracking and real-time quality prediction in complex industrial systems.

Authors 1

1. Emanuel Kashi Thienpont corresponding Aachen

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen