基于Transformer的出铝量决策算法研究与应用  

Research and application of aluminum output decision algorithm based on Transformer

作　　者：李明昊 李晋宏[1] Li Minghao;Li Jinhong(School of Information Science and Technology,North China University of Technology,Beijing 100144,China)

机构地区：[1]北方工业大学信息学院,北京100144

出　　处：《轻金属》2024年第4期19-24,共6页Light Metals

摘　　要：在传统电解铝工业生产中,铝电解槽的生产决策通常根据工艺技术人员多年的经验制定,其中出铝量是一个具有强耦合性的重要决策变量,其决策的好坏对电解槽的生产稳定性和产铝效率具有直接且重要的影响。本文提出了一种可深度挖掘数据频率特征和特征筛选的Transformer架构模型FDisformer,优化了传统的Transformer的encode编码层,进而挖掘数据趋势变化和更深层次的特征信息,同时引入了特征蒸馏模块,确保筛选出与出铝量决策这一任务强相关的特征。FDisformer在出铝量决策方面具有更高的性能指标,该模型的建立可以为后续的铝电解槽出铝量每日决策提供参考依据。In the traditional production of electrolytic aluminum industry,the production decisions of aluminum electrolytic cells are usually made based on the multi-years experience of process technicians.The aluminum output is an important decision variable with strong coupling,and the quality of its decision-making has a direct and important impact on the production stability and efficiency of the electrolytic cell.This paper proposes a Transformer architecture model FDisformer that can deeply mine data frequency features and feature filtering.It optimizes the encoding layer of traditional Transformer,thereby mining data trend changes and deeper feature information.At the same time,a feature distillation module is introduced to ensure the screening of features strongly related to the task of aluminum production decision-making.FDisformer has higher performance indicators in aluminum output decision-making,and the establishment of this model can provide a reference basis for the subsequent daily aluminum output decision-making in aluminum electrolytic cells.

关 键 词：铝电解槽 Transformer架构 频率特征挖掘 出铝量决策 特征蒸馏 

分 类 号：TF821[冶金工程—有色金属冶金]