| 引用本文: | 沈瑜,李江柽,梁栋,王若暄,吴红刚,马殷军,严源,马煜堃,刘佳英,闫佳荣.特征选择的高效时间序列预测模型[J].哈尔滨工业大学学报,2026,58(9):173.DOI:10.11918/202411025 |
| Shen Yu,Li Jiangcheng,Liang Dong,Wang Ruoxuan,Wu Honggang,Ma Yinjun,Yan Yuan,Ma Yukun,Liu Jiaying,Yan Jiarong.Efficient time series forecasting model for feature selection[J].Journal of Harbin Institute of Technology,2026,58(9):173.DOI:10.11918/202411025 |
|
| |
|
|
| 本文已被:浏览 693次 下载 57次 |
 码上扫一扫! |
|
|
| 特征选择的高效时间序列预测模型 |
|
沈瑜1,李江柽1,梁栋2,3,王若暄1,吴红刚2,3,马殷军4,严源2,3,马煜堃1,刘佳英1,闫佳荣1
|
|
(1.兰州交通大学 电子与信息工程学院,兰州 730070;2.中铁西北科学研究院有限公司,兰州 730046; 3.中国中铁科学研究院集团有限公司,成都 610031;4.中国铁路青藏集团有限公司,西宁 810007)
|
|
| 摘要: |
| 为解决时间序列预测方法普遍存在计算复杂度高、推理速度慢以及关键信息容易丢失的问题,提出一种特征选择的高效时间序列预测模型(Targetformer)。该模型采用Transformer仅编码器结构,并引入倒置输入编码块,将输入时间序列转换为独立的Token;提出特征选择性注意力机制,从映射后的时序Token中提取关键特征;最后利用投影层将提取的特征映射为最终的预测结果。 在6个数据集上的实验结果表明:相较于最优基线模型,Targetformer的均方误差和平均绝对误差分别降低5.7%和2.0%,均取得最优的预测性能,验证了其有效性和实用性。该模型为时间序列预测领域提供一种高效且实用的解决方案。 |
| 关键词: 时间序列预测 Targetformer 特征选择性注意力机制 倒置输入编码块 Transformer仅编码器结构 |
| DOI:10.11918/202411025 |
| 分类号:TP183 |
| 文献标识码:A |
| 基金项目:甘肃省技术创新引导计划企业研发机构能力建设专项资助(23CXJA0011);中国中铁股份有限公司科技研究开发计划(2022-重大专项-07) |
|
| Efficient time series forecasting model for feature selection |
|
Shen Yu1,Li Jiangcheng1,Liang Dong2,3,Wang Ruoxuan1,Wu Honggang2,3,Ma Yinjun4,Yan Yuan2,3,Ma Yukun1,Liu Jiaying1,Yan Jiarong1
|
|
(1.School of Electronic and Information Engineering, Lanzhou Jiaotong University, Lanzhou 730070, China; 2.Northwest Research Institute Co., Ltd. of C.R.E.C., Lanzhou 730046, China; 3.China Railway Academy Group Co., Ltd., Chengdu 610031, China; 4.China Railway Qinghai-Tibet Group Co., Ltd., Xining 810007, China)
|
| Abstract: |
| In order to solve the problems of high computational complexity, slow inference speed and easy loss of key information prevalent in time series prediction methods, an efficient time series prediction model with feature selection (Targetformer) is proposed. The model adopts the Transformer encoder-only structure and introduces an inverted embedding to convert the input time series into independent Token; A feature-selective attention mechanism is proposed to extract the key features from the mapped time-series Tokens. Finally, a projection layer is used to map the extracted features into the final prediction results. Experimental results on six datasets show that, compared with the best baseline model, Targetformer reduces mean squared error and mean absolute error by 5.7% and 2.0%, respectively, and achieves the best prediction performance in terms of both metrics, verifying its effectiveness and practical applicability. The model provides an efficient and practical solution in the field of time series prediction. |
| Key words: time series forecasting Targetformer feature-selective attention mechanism inverted embedding Transformer encoder-only structure |
|
|
|
|