Efficient Prediction of Spatiotemporal Temperature Distribution Caused by Carbon Combustion in a Packed Bed Using Long Short-Term Memory Networks

Sintering is a crucial agglomeration process in modern steel production, and the temperature distribution within the packed bed is essential for reducing energy consumption and improving sinter quality. Traditional approaches, such as sintering pot experiments and numerical simulations, offer valuable insights but are hindered by their labor-intensive nature and high computational demands, respectively. These challenges limit the ability to provide timely and quantitative feedback for practical on-site production. To overcome these limitations, a previously developed two-dimensional computational fluid dynamics (CFD)-based model of carbon combustion in a packed bed is used to simulate spatiotemporal temperature distribution under 625 different sintering conditions. These simulations generate a dataset of 12,500 spatiotemporal temperature series based on bed height and sintering time. Building on this dataset, a data-driven model, which based on Long Short-Term Memory (LSTM) networks, is proposed to predict the spatiotemporal distribution of temperatures. The model employs a single-layer LSTM network with 15 hidden units to capture the temperature series over time, followed by sampling at various bed heights to obtain spatial distribution. The results demonstrate that, compared to CFD simulations, the data-driven model enhances prediction efficiency by at least three orders of magnitude, achieving an absolute error of 9.77 K and a mean relative error of 1.38 pct. This data-driven model offers a highly efficient method for predicting spatiotemporal temperature distribtuions, presenting significant implications for the intelligent production and control of sintering processes.