MLDF-Net: Multivariate Learnable Decomposition Forecasting Network for Industrial Predictive Modeling

Accurate multi-step forecasting of key state variables is critical for predictive maintenance and optimization control in industrial processes. Transformer-based models have emerged as prominent solutions owing to their capability to dynamically capture long-term temporal dependencies through positional encoding and attention mechanisms. However, these models often neglect multivariate correlations between process variables with different physical meanings and face difficulties in handling multi-scale temporal patterns. In this paper, a novel multivariate learnable decomposition forecasting network for predictive modeling is proposed in industrial processes. Specifically, a learnable decomposition module is employed to dynamically model trend information across multiple scales. For each decomposed component, a dedicated prediction network is tailored based on real-world industrial process characteristics. The variate tokens of iTransformer are utilized to preserve the physical semantics of variables, while multi-head attention is leveraged to capture cross-variable dependencies. Finally, the superiority of the proposed framework is demonstrated through a real-world application in the alumina dissolution process.