基于LSTM-Attention的卷烟产能预测研究

Research on Industrial Productivity Prediction Based on Dynamic Attention Mechanism

摘要: 针对智能制造产能预测中数据层级复杂的问题，采用某卷烟厂四年的多层级生产数据，经系统化预处理后确定最优建模粒度，并设计了基于班次分组的线性插值重构方法解决时序不连续性问题。在此基础上，提出了融合长短期记忆(long short-term memory, LSTM)与注意力机制的轻量化预测框架，并引入早停策略防止过拟合。实验结果表明，模型在预测准确性和稳定性方面优于对比方法，远超传统的经验计算方式。同时深入分析了多步预测误差累积、泛化能力下降等固有局限性的根本原因，并提出了系统性的改进策略。结论表明，LSTM-Attention融合框架在保证预测精度的同时，具备良好的工程适用性，为智能制造环境下的生产优化决策提供了重要参考。

Abstract: To address data hierarchy complexity and computational redundancy in smart manufacturing capacity prediction, multi-level production data from a cigarette factory is preprocessed to determine optimal modeling granularity. A shift-grouping method with linear interpolation resolves time-series discontinuity. A lightweight forecasting framework integrating Long Short-Term Memory (LSTM) and an attention mechanism are proposed, incorporating early stopping to prevent overfitting. Results demonstrate superior prediction accuracy and stability over benchmarks, significantly outperforming traditional empirical methods. Fundamental causes of limitations—including multi-step error accumulation and reduced generalization—were analyzed. Systematic improvement strategies are formulated based on this analysis. The LSTM-Attention framework ensures prediction precision while offering strong engineering applicability, providing vital reference for production optimization in smart manufacturing.