FECAM: Frequency enhanced channel attention mechanism for time series forecasting

架空(工程) 频道(广播) 离散余弦变换 噪音(视频) 算法 离散傅里叶变换(通用) 计算机科学 快速傅里叶变换 人工智能 吉布斯现象 转化(遗传学) 频域 傅里叶变换 电信 机器学习 数学 傅里叶分析 短时傅里叶变换 图像(数学) 计算机视觉 数学分析 操作系统 基因 化学 生物化学
作者
Maowei Jiang,Pengyu Zeng,Kai Wang,Huan Liu,Wenbo Chen,Haoran Liu
出处
期刊:Advanced Engineering Informatics [Elsevier BV]
卷期号:58: 102158-102158 被引量:173
标识
DOI:10.1016/j.aei.2023.102158
摘要

Time series forecasting (TSF) is a challenging problem in various real-world scenarios, such as industry, energy, weather, traffic, economics, and earthquake warning. TSF demands the model to have a high prediction accuracy. Despite the promising performance of deep learning-based methods in TSF tasks, mainstream forecasting models may sometimes produce results that deviate from the actual ground truth. Our analysis suggests that this may be attributed to the models’ limited ability to capture the frequency information that is abundantly present in real-world datasets. Currently, the Fourier Transform (FT) is the most widely used method for extracting frequency information, but it has some issues that lead to poor model performance, such as high-frequency noise caused by the Gibbs phenomenon and computational overhead of the inverse transformation in the FT-IFT process. To address these issues, we propose a novel frequency enhanced channel attention mechanism (FECAM) that models frequency interdependencies between channels based on Discrete Cosine Transform (DCT), which inherently mitigates the high-frequency noise caused by problematic periodicity during Fourier Transform. This approach improves the model’s capability to extract frequency features and resolves computational overhead concerns that arise from inverse transformations. Our contributions are threefold: (1) We propose a novel frequency enhanced channel attention mechanism that models frequency interdependencies between channels based on DCT, which improves the model’s capability to extract frequency features and resolves computational overhead concerns that arise from inverse transformations; (2) We theoretically prove that our method mitigates the Gibbs phenomenon, which introduces high frequency noise during Fourier Transform. We demonstrate that the result of 1D GAP linearly varies with the lowest frequency component of 1D DCT; (3) We demonstrate the generalization ability of the proposed method FECAM by embedding it into other networks, resulting in significant performance improvements when compared to the original model, with only a minor increase in parameters. Furthermore, we conduct extensive experiments on six different real-world TSF datasets to validate the effectiveness of our proposed model and compare it with several existing state-of-the-art models. Our findings indicate that the FECAM model is superior to these models in terms of accuracy, making it a promising solution for TSF in diverse real-world scenarios. Our codes and data are available at https://github.com/Zero-coder/FECAM.
最长约 10秒,即可获得该文献文件

科研通智能强力驱动
Strongly Powered by AbleSci AI
科研通是完全免费的文献互助平台,具备全网最快的应助速度,最高的求助完成率。 对每一个文献求助,科研通都将尽心尽力,给求助人一个满意的交代。
实时播报
ggdh发布了新的文献求助10
刚刚
1秒前
余其钵发布了新的文献求助10
1秒前
酷炫的流沙完成签到,获得积分10
1秒前
1秒前
1秒前
3秒前
4秒前
布丁发布了新的文献求助10
4秒前
4秒前
5秒前
海大彭于晏完成签到,获得积分10
5秒前
ZJB关闭了ZJB的文献求助
5秒前
5秒前
花卷发布了新的文献求助10
5秒前
5秒前
李小白的应助被jxxyzm采纳,获得10
6秒前
6秒前
7秒前
lh完成签到,获得积分10
8秒前
8秒前
8秒前
科目三的应助被执着的立果采纳,获得10
9秒前
Kerrie完成签到,获得积分10
9秒前
10秒前
佳洁宝贝发布了新的文献求助10
10秒前
Yuqi发布了新的文献求助10
11秒前
许愿完成签到 ,获得积分10
11秒前
华仔的应助被AKYDXS采纳,获得10
12秒前
阿梨完成签到,获得积分20
12秒前
云舒发布了新的文献求助10
12秒前
饼大王完成签到,获得积分10
12秒前
真实的麦片完成签到,获得积分10
13秒前
13秒前
cwj的应助被化学新人采纳,获得10
13秒前
MYR完成签到,获得积分10
13秒前
俊逸夜蓉发布了新的文献求助10
14秒前
许愿关注了科研通微信公众号
14秒前
14秒前
CC完成签到,获得积分10
14秒前
高分求助中
(应助此贴封号)【重要!!请各用户(尤其是新用户)详细阅读】【科研通的精品贴汇总】 10000
拉伸高分子熔体中缠结行为的模拟和小角中子散射研究 1700
The World Atlas of Wine 9th edition 1000
奈特基础组织学:配套组织病理学(第4版)(Netter's Essential Histology With Correlated Histopathology) 800
John McDowells Bildungstheorie 800
八旗制度の研究 600
plantTFDB 508
热门求助领域 (近24小时)
化学 材料科学 医学 生物 纳米技术 计算机科学 工程类 有机化学 化学工程 生物化学 物理 复合材料 内科学 催化作用 光电子学 人工智能 细胞生物学 无机化学 物理化学 数学
热门帖子
关注 科研通微信公众号,转发送积分 8168423
求助须知:如何正确求助?哪些是违规求助? 9561356
关于积分的说明 21761245
捐赠科研通 7759810
什么是DOI,文献DOI怎么找? 3429998
关于科研通互助平台的介绍 2553872
邀请新用户注册赠送积分活动 2474665