A Dual-Channel Multimodal Data Fusion Approach for Hot Metal Temperature Prediction in Blast Furnaces

Predicting the thermal state of the blast furnace is a crucial step in stabilizing operations, improving process efficiency, and minimizing disturbances that can increase energy consumption and emissions. Traditional first-principles models are limited by the system’s complexity, involving temporally and spatially distributed variables and non-ideal conditions. Recent advances in machine learning enable the development of predictive models using diverse data types. This study presents the Dual-Channel Fusion Analysis Network (DCFANet), a deep learning model specifically designed for this task. DCFANet processes multi-modal datasets in parallel, combining a multi-layer convolutional neural network to extract spatiotemporal features from tuyere images with a Gated Recurrent Unit, equipped with an attention mechanism, to extract time-series features from numerical data. These features are then integrated to predict future hot metal temperatures. Experiments using four months of industrial production data demonstrate that DCFANet achieves a mean absolute error of 5.4 °C and a correlation coefficient of 0.93 between one-step ahead predictions and actual values. The results highlight DCFANet as a versatile tool for blast furnace operation guidance.