《计算机应用》唯一官方网站 ›› 2025, Vol. 45 ›› Issue (12): 3855-3863.DOI: 10.11772/j.issn.1001-9081.2024121818
朱昶胜1, 杨琛1, 冯文芳2, 袁培文1
收稿日期:2024-12-27
修回日期:2025-03-14
接受日期:2025-03-18
发布日期:2025-03-27
出版日期:2025-12-10
通讯作者:
杨琛
作者简介:朱昶胜(1972—),男,甘肃秦安人,教授,博士,主要研究方向:高性能计算与大数据、制造业信息化系统与工程基金资助:Changsheng ZHU1, Chen YANG1, Wenfang FENG2, Peiwen YUAN1
Received:2024-12-27
Revised:2025-03-14
Accepted:2025-03-18
Online:2025-03-27
Published:2025-12-10
Contact:
Chen YANG
About author:ZHU Changsheng, born in 1972, Ph. D., professor. His research interests include high-performance computing and big data, information systems and engineering in manufacturing.Supported by:摘要:
简单线性模型的时间序列预测质量通常超过Transformer等深度模型;而在具有大量通道的数据集上,深度模型尤其是多层感知器(MLP)的性能反而可超过简单线性模型。针对简单线性模型和MLP在时间序列预测中的误差功率谱差异,提出一种基于MLP的高频增强型时间序列预测模型HiFNet(High-Frequency Network)。首先,利用MLP在低频段的拟合能力;其次,通过自适应序列分解(ASD)模块及分组线性层解决MLP高频段易过拟合以及通道独立策略不能有效应对通道冗余的问题,从而增强MLP在高频段的鲁棒性;最后,对HiFNet在气象、电力和交通等领域的标准数据集上进行实验。结果表明:HiFNet的均方误差(MSE)在最佳情况下相较于NLinear、RLinear、SegRNN(Segment Recurrent Neural Network)和PatchTST(Patch Time Series Transformer)分别降低了23.6%、10.0%、35.1%和6.5%,而分组线性层通过学习通道相关性的低秩表达减轻了通道冗余的影响。
中图分类号:
朱昶胜, 杨琛, 冯文芳, 袁培文. 基于多层感知器的高频增强型时间序列预测模型[J]. 计算机应用, 2025, 45(12): 3855-3863.
Changsheng ZHU, Chen YANG, Wenfang FENG, Peiwen YUAN. High-frequency enhanced time series prediction model based on multi-layer perceptron[J]. Journal of Computer Applications, 2025, 45(12): 3855-3863.
| 数据集 | 通道数 | 时间步 | 细粒度/min | 数据划分 |
|---|---|---|---|---|
| Weather | 21 | 52 696 | 10 | 7∶2∶1 |
| ECL | 321 | 26 304 | 60 | 7∶2∶1 |
| Traffic | 862 | 17 545 | 60 | 7∶2∶1 |
| Solar Energy | 137 | 51 938 | 10 | 7∶2∶1 |
表1 实验数据集
Tab. 1 Experimental datasets
| 数据集 | 通道数 | 时间步 | 细粒度/min | 数据划分 |
|---|---|---|---|---|
| Weather | 21 | 52 696 | 10 | 7∶2∶1 |
| ECL | 321 | 26 304 | 60 | 7∶2∶1 |
| Traffic | 862 | 17 545 | 60 | 7∶2∶1 |
| Solar Energy | 137 | 51 938 | 10 | 7∶2∶1 |
| 数据集 | 预测长度 | HiFNet | NLinear | RLinear | SegRNN | PatchTST | |||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| MSE | MAE | MSE | MAE | MSE | MAE | MSE | MAE | MSE | MAE | ||
| Weather | 96 | 0.144 | 0.193 | 0.145 | 0.194 | 0.145 | 0.193 | 0.147 | 0.208 | 0.150 | 0.199 |
| 192 | 0.185 | 0.232 | 0.189 | 0.237 | 0.189 | 0.235 | 0.193 | 0.256 | 0.198 | 0.243 | |
| 336 | 0.237 | 0.272 | 0.241 | 0.279 | 0.240 | 0.274 | 0.244 | 0.297 | 0.253 | 0.282 | |
| 720 | 0.313 | 0.325 | 0.317 | 0.334 | 0.316 | 0.328 | 0.327 | 0.356 | 0.327 | 0.342 | |
| ECL | 96 | 0.128 | 0.223 | 0.135 | 0.238 | 0.135 | 0.231 | 0.133 | 0.231 | 0.131 | 0.227 |
| 192 | 0.146 | 0.238 | 0.150 | 0.243 | 0.151 | 0.245 | 0.151 | 0.250 | 0.150 | 0.243 | |
| 336 | 0.163 | 0.255 | 0.167 | 0.260 | 0.167 | 0.261 | 0.170 | 0.270 | 0.185 | 0.288 | |
| 720 | 0.198 | 0.287 | 0.206 | 0.295 | 0.206 | 0.293 | 0.207 | 0.305 | 0.218 | 0.311 | |
| Traffic | 96 | 0.377 | 0.261 | 0.424 | 0.295 | 0.419 | 0.291 | 0.555 | 0.274 | 0.384 | 0.267 |
| 192 | 0.393 | 0.269 | 0.457 | 0.299 | 0.430 | 0.295 | 0.576 | 0.275 | 0.403 | 0.274 | |
| 336 | 0.409 | 0.277 | 0.482 | 0.306 | 0.444 | 0.302 | 0.671 | 0.280 | 0.415 | 0.283 | |
| 720 | 0.440 | 0.295 | 0.576 | 0.325 | 0.472 | 0.318 | 0.679 | 0.289 | 0.452 | 0.299 | |
| Solar Energy | 96 | 0.195 | 0.239 | 0.230 | 0.266 | 0.231 | 0.264 | 0.197 | 0.250 | 0.198 | 0.248 |
| 192 | 0.185 | 0.252 | 0.259 | 0.281 | 0.259 | 0.278 | 0.187 | 0.257 | 0.199 | 0.254 | |
| 336 | 0.196 | 0.260 | 0.279 | 0.288 | 0.280 | 0.290 | 0.197 | 0.262 | 0.218 | 0.274 | |
| 720 | 0.224 | 0.266 | 0.281 | 0.289 | 0.280 | 0.289 | 0.225 | 0.266 | 0.238 | 0.285 | |
表2 不同模型在4个数据集上不同预测长度的指标对比
Tab. 2 Comparison of indicators with different prediction lengths for different models on four datasets
| 数据集 | 预测长度 | HiFNet | NLinear | RLinear | SegRNN | PatchTST | |||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| MSE | MAE | MSE | MAE | MSE | MAE | MSE | MAE | MSE | MAE | ||
| Weather | 96 | 0.144 | 0.193 | 0.145 | 0.194 | 0.145 | 0.193 | 0.147 | 0.208 | 0.150 | 0.199 |
| 192 | 0.185 | 0.232 | 0.189 | 0.237 | 0.189 | 0.235 | 0.193 | 0.256 | 0.198 | 0.243 | |
| 336 | 0.237 | 0.272 | 0.241 | 0.279 | 0.240 | 0.274 | 0.244 | 0.297 | 0.253 | 0.282 | |
| 720 | 0.313 | 0.325 | 0.317 | 0.334 | 0.316 | 0.328 | 0.327 | 0.356 | 0.327 | 0.342 | |
| ECL | 96 | 0.128 | 0.223 | 0.135 | 0.238 | 0.135 | 0.231 | 0.133 | 0.231 | 0.131 | 0.227 |
| 192 | 0.146 | 0.238 | 0.150 | 0.243 | 0.151 | 0.245 | 0.151 | 0.250 | 0.150 | 0.243 | |
| 336 | 0.163 | 0.255 | 0.167 | 0.260 | 0.167 | 0.261 | 0.170 | 0.270 | 0.185 | 0.288 | |
| 720 | 0.198 | 0.287 | 0.206 | 0.295 | 0.206 | 0.293 | 0.207 | 0.305 | 0.218 | 0.311 | |
| Traffic | 96 | 0.377 | 0.261 | 0.424 | 0.295 | 0.419 | 0.291 | 0.555 | 0.274 | 0.384 | 0.267 |
| 192 | 0.393 | 0.269 | 0.457 | 0.299 | 0.430 | 0.295 | 0.576 | 0.275 | 0.403 | 0.274 | |
| 336 | 0.409 | 0.277 | 0.482 | 0.306 | 0.444 | 0.302 | 0.671 | 0.280 | 0.415 | 0.283 | |
| 720 | 0.440 | 0.295 | 0.576 | 0.325 | 0.472 | 0.318 | 0.679 | 0.289 | 0.452 | 0.299 | |
| Solar Energy | 96 | 0.195 | 0.239 | 0.230 | 0.266 | 0.231 | 0.264 | 0.197 | 0.250 | 0.198 | 0.248 |
| 192 | 0.185 | 0.252 | 0.259 | 0.281 | 0.259 | 0.278 | 0.187 | 0.257 | 0.199 | 0.254 | |
| 336 | 0.196 | 0.260 | 0.279 | 0.288 | 0.280 | 0.290 | 0.197 | 0.262 | 0.218 | 0.274 | |
| 720 | 0.224 | 0.266 | 0.281 | 0.289 | 0.280 | 0.289 | 0.225 | 0.266 | 0.238 | 0.285 | |
| 数据集 | 预测长度 | HiFNet | HiFNet-AvgPool | ||
|---|---|---|---|---|---|
| MSE | MAE | MSE | MAE | ||
| Weather | 96 | 0.144 | 0.193 | 0.144 | 0.192 |
| 192 | 0.185 | 0.232 | 0.187 | 0.235 | |
| 336 | 0.237 | 0.272 | 0.239 | 0.275 | |
| 720 | 0.313 | 0.325 | 0.315 | 0.326 | |
| ECL | 96 | 0.128 | 0.223 | 0.130 | 0.225 |
| 192 | 0.146 | 0.238 | 0.148 | 0.241 | |
| 336 | 0.163 | 0.255 | 0.167 | 0.260 | |
| 720 | 0.198 | 0.287 | 0.202 | 0.289 | |
| Traffic | 96 | 0.377 | 0.261 | 0.383 | 0.265 |
| 192 | 0.393 | 0.269 | 0.402 | 0.274 | |
| 336 | 0.409 | 0.277 | 0.417 | 0.280 | |
| 720 | 0.440 | 0.295 | 0.442 | 0.301 | |
表3 替换ASD模块的消融实验结果
Tab. 3 Ablation study results of replacing ASD module
| 数据集 | 预测长度 | HiFNet | HiFNet-AvgPool | ||
|---|---|---|---|---|---|
| MSE | MAE | MSE | MAE | ||
| Weather | 96 | 0.144 | 0.193 | 0.144 | 0.192 |
| 192 | 0.185 | 0.232 | 0.187 | 0.235 | |
| 336 | 0.237 | 0.272 | 0.239 | 0.275 | |
| 720 | 0.313 | 0.325 | 0.315 | 0.326 | |
| ECL | 96 | 0.128 | 0.223 | 0.130 | 0.225 |
| 192 | 0.146 | 0.238 | 0.148 | 0.241 | |
| 336 | 0.163 | 0.255 | 0.167 | 0.260 | |
| 720 | 0.198 | 0.287 | 0.202 | 0.289 | |
| Traffic | 96 | 0.377 | 0.261 | 0.383 | 0.265 |
| 192 | 0.393 | 0.269 | 0.402 | 0.274 | |
| 336 | 0.409 | 0.277 | 0.417 | 0.280 | |
| 720 | 0.440 | 0.295 | 0.442 | 0.301 | |
| 数据集 | 预测长度 | HiFNet | MLP | ||
|---|---|---|---|---|---|
| MSE | MAE | MSE | MAE | ||
| Weather | 96 | 0.144 | 0.193 | 0.146 | 0.207 |
| 192 | 0.185 | 0.232 | 0.193 | 0.256 | |
| 336 | 0.237 | 0.272 | 0.245 | 0.296 | |
| 720 | 0.313 | 0.325 | 0.325 | 0.355 | |
| ECL | 96 | 0.128 | 0.223 | 0.131 | 0.225 |
| 192 | 0.146 | 0.238 | 0.148 | 0.240 | |
| 336 | 0.163 | 0.255 | 0.165 | 0.258 | |
| 720 | 0.198 | 0.287 | 0.201 | 0.289 | |
| Traffic | 96 | 0.377 | 0.261 | 0.380 | 0.262 |
| 192 | 0.393 | 0.269 | 0.400 | 0.271 | |
| 336 | 0.409 | 0.277 | 0.413 | 0.278 | |
| 720 | 0.440 | 0.295 | 0.443 | 0.298 | |
表4 移除ASD模块和分组线性层的消融实验结果
Tab. 4 Ablation study results of removing ASD module and grouped linear layer
| 数据集 | 预测长度 | HiFNet | MLP | ||
|---|---|---|---|---|---|
| MSE | MAE | MSE | MAE | ||
| Weather | 96 | 0.144 | 0.193 | 0.146 | 0.207 |
| 192 | 0.185 | 0.232 | 0.193 | 0.256 | |
| 336 | 0.237 | 0.272 | 0.245 | 0.296 | |
| 720 | 0.313 | 0.325 | 0.325 | 0.355 | |
| ECL | 96 | 0.128 | 0.223 | 0.131 | 0.225 |
| 192 | 0.146 | 0.238 | 0.148 | 0.240 | |
| 336 | 0.163 | 0.255 | 0.165 | 0.258 | |
| 720 | 0.198 | 0.287 | 0.201 | 0.289 | |
| Traffic | 96 | 0.377 | 0.261 | 0.380 | 0.262 |
| 192 | 0.393 | 0.269 | 0.400 | 0.271 | |
| 336 | 0.409 | 0.277 | 0.413 | 0.278 | |
| 720 | 0.440 | 0.295 | 0.443 | 0.298 | |
| 模型 | 训练时间/s | 浮点运算次数/MFLOPs |
|---|---|---|
| HiFNet | 100 | 4 052 |
| MLP | 90 | 225 |
| NLinear | 178 | 140 |
| RLinear | 240 | 140 |
| SegRNN | 100 | 7 116 |
| PatchTST | 847 | 198 873 |
表5 不同模型训练时间与浮点运算次数的对比
Tab. 5 Comparison of training time and FLOPs of different models
| 模型 | 训练时间/s | 浮点运算次数/MFLOPs |
|---|---|---|
| HiFNet | 100 | 4 052 |
| MLP | 90 | 225 |
| NLinear | 178 | 140 |
| RLinear | 240 | 140 |
| SegRNN | 100 | 7 116 |
| PatchTST | 847 | 198 873 |
| [1] | GUO H, ZHANG J, ZHANG J, et al. Prediction of highway blocking loss based on ensemble learning fusion model[J]. Electronics, 2022, 11(17): No.2792. |
| [2] | YU W, WANG S, ZHANG C, et al. Integrating spatio-temporal and generative adversarial networks for enhanced nowcasting performance[J]. Remote Sensing, 2023, 15(15): No.3720. |
| [3] | 孙卓远,吕学文,王继军. 基于门控多层感知机和Informer的多通道电力负荷预测[J]. 人工智能与机器人研究, 2024, 13(2): 375-387. |
| SUN Z Y, LYU X W, WANG J J. Multi-channel power load forecasting based on gated multilayer perceptron and Informer[J]. Artificial Intelligence and Robotics Research, 2024, 13(2): 375-387. | |
| [4] | ZHOU H, ZHANG S, PENG J, et al. Informer: beyond efficient Transformer for long sequence time-series forecasting[C]// Proceedings of the 35th AAAI Conference on Artificial Intelligence. Palo Alto: AAAI Press, 2021: 11106-11115. |
| [5] | ZENG A, CHEN M, ZHANG L, et al. Are Transformers effective for time series forecasting?[C]// Proceedings of the 37th AAAI Conference on Artificial Intelligence. Palo Alto: AAAI Press, 2023:11121-11128. |
| [6] | LI W, MENG X, CHEN C, et al. Mlinear: rethink the linear model for time-series forecasting [EB/OL]. [2024-11-01]. . |
| [7] | LI Z, QI S, LI Y, et al. Revisiting long-term time series forecasting: an investigation on linear mapping [EB/OL]. [2024-11-01]. . |
| [8] | LAKSHMI K, REDDY D K, SHARMA M R. Integration of hybrid ARIMA artificial neural networks for accurate platinum price prediction[J]. Communications on Applied Nonlinear Analysis, 2024, 31(6s): 61-73. |
| [9] | HWANG Y, TONG A, CHOI J. Automatic construction of nonparametric relational regression models for multiple time series[C]// Proceedings of the 33rd International Conference on Machine Learning. New York: JMLR.org, 2016: 3030-3039. |
| [10] | FU Z, WU Y, LIU X. A tensor-based deep LSTM forecasting model capturing the intrinsic connection in multivariate time series[J]. Applied Intelligence, 2023, 53(12): 15873-15888. |
| [11] | LIN S, LIN W, WU W, et al. SegRNN: segment recurrent neural network for long-term time series forecasting [EB/OL]. [2024-11-01]. . |
| [12] | DARLOW L, DENG Q, HASSAN A, et al. DAM: towards a foundation model for time series forecasting [EB/OL]. [2024-11-01]. . |
| [13] | NIE Y, NGUYEN N H, SINTHONG P, et al. A time series is worth 64 words: long-term forecasting with Transformers [EB/OL]. [2024-11-01]. . |
| [14] | LI Z, ZHANG X, DONG Z. TSF-Transformer: a time series forecasting model for exhaust gas emission using Transformer[J]. Applied Intelligence, 2023, 53(13): 17211-17225. |
| [15] | ZHOU T, MA Z, WEN Q, et al. FEDformer: frequency enhanced decomposed transformer for long-term series forecasting[C]// Proceedings of the 39th International Conference on Machine Learning. New York: JMLR.org, 2022: 27268-27286. |
| [16] | ZHANG C, ZHOU T, WEN Q, et al. TFAD: a decomposition time series anomaly detection architecture with time-frequency analysis[C]// Proceedings of the 31st ACM International Conference on Information and Knowledge Management. New York.: ACM, 2022: 2497-2507. |
| [17] | XU Z Q J, ZHANG Y, XIAO Y. Training behavior of deep neural network in frequency domain[C]// Proceedings of the 2019 International Conference on Neural Information Processing, LNCS 11953. Cham: Springer, 2019: 264-274. |
| [18] | RAHANMAN N, BARATIN A, ARPIT D, et al. On the spectral bias of neural networks[C]// Proceedings of the 36th International Conference on Machine Learning. New York: JMLR.org, 2019: 5301-5310. |
| [19] | XU Z Q J, ZHANG Y, LUO T, et al. Frequency principle: Fourier analysis sheds light on deep neural networks[J]. Computer Physics Communications, 2019, 28(5): 1746-1767. |
| [20] | XU Z, ZENG A, XU Q. FITS: modeling time series with 10k parameters[EB/OL]. [2024-11-01].. |
| [21] | EKAMBARAM V, JATI A, NGUYEN N, et al. TSMixer: lightweight MLP-Mixer model for multivariate time series forecasting[C]// Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. New York: ACM, 2023: 459-469. |
| [22] | RAGHU M, POOLE B, KLEINBERG J, et al. On the expressive power of deep neural networks[C]// Proceedings of the 34th International Conference on Machine Learning. New York: JMLR.org, 2017: 2847-2854. |
| [23] | POOLE B, LAHIRI S, RAGHU M, et al. Exponential expressivity in deep neural networks through transient chaos[C]// Proceedings of the 30th International Conference on Neural Information Processing Systems. Red Hook: Curran Associates Inc., 2016: 3368-3376. |
| [24] | KIM T, KIM J, TAE Y, et al. Reversible instance normalization for accurate time-series forecasting against distribution shift[EB/OL]. [2024-11-01]. . |
| [25] | RAMACHANDRAN P, ZOPH B, LE Q V. Searching for activation functions[EB/OL]. [2024-11-01]. . |
| [1] | 古松健, 吴福祥, 高向阳, 杨梦杰, 詹忆冰, 程俊. 基于几何模型预测控制的移动机器人轨迹跟踪算法[J]. 《计算机应用》唯一官方网站, 2025, 45(9): 3026-3035. |
| [2] | 王翔, 陈志祥, 毛国君. 融合局部和全局相关性的多变量时间序列预测方法[J]. 《计算机应用》唯一官方网站, 2025, 45(9): 2806-2816. |
| [3] | 王慧斌, 胡展傲, 胡节, 徐袁伟, 文博. 基于分段注意力机制的时间序列预测模型[J]. 《计算机应用》唯一官方网站, 2025, 45(7): 2262-2268. |
| [4] | 郭茂祖, 崔正, 赵玲玲, 张庆宇. 基于频域注意力的结构地震响应预测方法[J]. 《计算机应用》唯一官方网站, 2025, 45(11): 3730-3738. |
| [5] | 任烈弘, 黄铝文, 田旭, 段飞. 基于DFT的频率敏感双分支Transformer多变量长时间序列预测方法[J]. 《计算机应用》唯一官方网站, 2024, 44(9): 2739-2746. |
| [6] | 范黎林, 曹富康, 王琬婷, 杨凯, 宋钊瑜. 基于需求模式自适应匹配的间歇性需求预测方法[J]. 《计算机应用》唯一官方网站, 2024, 44(9): 2747-2755. |
| [7] | 徐泽鑫, 杨磊, 李康顺. 较短的长序列时间序列预测模型[J]. 《计算机应用》唯一官方网站, 2024, 44(6): 1824-1831. |
| [8] | 曾渝, 张洋, 曾尚, 付茂栗, 何启学, 曾林隆. 基于多尺度门控膨胀卷积网络的时间序列预测算法[J]. 《计算机应用》唯一官方网站, 2024, 44(11): 3427-3434. |
| [9] | 刘文博, 于连飞, 谢冬梅, 蔡闯, 曲志坚, 任崇广. 基于多尺度特征融合的时间序列长期预测模型[J]. 《计算机应用》唯一官方网站, 2024, 44(11): 3435-3441. |
| [10] | 范艺扬, 张洋, 曾尚, 曾渝, 付茂栗. 基于分解和频域特征提取的多变量长时间序列预测模型[J]. 《计算机应用》唯一官方网站, 2024, 44(11): 3442-3448. |
| [11] | 夏进, 王正群, 朱世明. 基于时间序列分解的交通流量预测模型[J]. 《计算机应用》唯一官方网站, 2023, 43(4): 1129-1135. |
| [12] | 郎祎平, 毛文涛, 罗铁军, 范黎林, 任颖莹, 刘侠. 间歇性时间序列的可预测性评估及联合预测方法[J]. 《计算机应用》唯一官方网站, 2022, 42(9): 2722-2731. |
| [13] | 吴明晖, 张广洁, 金苍宏. 基于多模态信息融合的时间序列预测模型[J]. 《计算机应用》唯一官方网站, 2022, 42(8): 2326-2332. |
| [14] | 王艺霏, 于雷, 滕飞, 宋佳玉, 袁玥. 基于长-短时序特征融合的资源负载预测模型[J]. 《计算机应用》唯一官方网站, 2022, 42(5): 1508-1515. |
| [15] | 胡鹤轩, 隋华超, 胡强, 张晔, 胡震云, 马能武. 基于图注意力网络与双阶注意力机制的径流预报模型[J]. 《计算机应用》唯一官方网站, 2022, 42(5): 1607-1615. |
| 阅读次数 | ||||||
|
全文 |
|
|||||
|
摘要 |
|
|||||