A Long-Term Interval Prediction Method for Industrial Oxygen Demand Using Granulation-Based Attention Mechanism and Multiscale Decomposition

Abstract Accurate long-term interval forecasting of downstream oxygen demand is essential for supporting scheduling engineers in large steel enterprises to make well-informed decisions about the oxygen supply network. This article introduces a new hybrid interval prediction approach that integrates multiscale decomposition with a custom-designed granulation-based transformer model. The multiscale decomposition is utilized to address the data's multifrequency components and high-frequency noise. Compared with the classical attention mechanisms, the proposed granulation-based attention mechanism helps the model better utilize the production semantics contained in the data. In contrast to traditional interval prediction methods that use non-differentiable loss functions, this article introduces a new differentiable loss function to address this limitation. This enables the optimization of model parameters without relying on heuristic optimization algorithms, simplifying the model training process. The industrial case study validates the effectiveness of the proposed approach, demonstrating better performance than both current state-of-the-art and newly developed methods.