EnvFormer: A Decomposition-Based Transformer for Multistep Burn-Through Point Prediction in the Sintering Process

Iron ore sintering is the process of preparing sinter for blast furnaces and is a key factor in determining ironmaking efficiency. During the sintering process, the real-time control of the burn-through point (BTP) determines the quality of the sinter. However, strong nonlinearity, time delay, and varying operating conditions prevent the model from providing accurate multistep predictions of the BTP. To overcome the above challenges, a decomposition-based encoder-decoder model, EnvFormer, is proposed in this article. As a variant of Transformer-based time-series forecasting models, EnvFormer captures complex nonlinear relationships and long-term dependencies in data. It deploys an iterative envelope decomposition (EnvD) module to simplify the complex raw signal into discernible components. The effectiveness of the proposed EnvFormer is verified through its application in a real industrial sintering process, and the results demonstrate that EnvFormer has better prediction accuracy for ten-step BTP prediction than state-of-the-art methods.