中国烟草学报2026，Vol.32Issue(3)：85-96,12.DOI:10.16472/j.chinatobacco.2025.T0397
基于频域门控Transformer的润叶过程参数预测方法
Parameter prediction of the tobacco leaf conditioning process based on a frequency-domain gated Transformer
摘要
Abstract
[Objective]To address the challenges of multivariate interactions,non-stationarity,and time delays in predicting outlet moisture during the hot air leaf conditioning process,this paper proposes ADGformer,a prediction model based on an adaptive frequency domain gated Transformer.[Methods]This model utilizes the fast Fourier transform to capture the periodic components of the time series.It incorporates an attention mechanism enhanced by frequency gating and introduces an adaptive frequency selection mechanism to extract key frequencies and filter out noise.[Results]The model achieves average relative reductions of 7.30%in mean squared error(MSE)and 7.26%in mean absolute error(MAE)across four prediction horizons,demonstrating superior robustness in long-term forecasting tasks.Furthermore,the single-sample inference latency is only 1.9 ms and the industrial qualification rate reaches 92.8%,which indicates a significant accuracy improvement in moisture prediction for the hot air leaf conditioning process.[Conclusion]The proposed ADGformer model provides an effective solution for the fine-grained control of the tobacco leaf conditioning process and offers new theoretical support for complex time-series prediction problems.关键词
烟叶水分预测/润叶过程/Transformer网络/频域分析/注意力机制Key words
tobacco leaf moisture prediction/leaf conditioning process/Transformer network/frequency-domain analysis/attention mechanism引用本文复制引用
王先兵,龚剑,黄毅,涂宸宇,吴忠泽,战思聪,龙军..基于频域门控Transformer的润叶过程参数预测方法[J].中国烟草学报,2026,32(3):85-96,12.基金项目
湖南中烟工业有限责任公司智能制造科研重大专项项目"制丝线全过程水分链模型构建及精准技术研究"(KY2025ZB0008) （KY2025ZB0008）
