Skip to main content
QUICK REVIEW

[论文解读] Multi-resolution Time-Series Transformer for Long-term Forecasting

Yitian Zhang, Liheng Ma|arXiv (Cornell University)|Nov 7, 2023
Time Series Analysis and Forecasting被引用 18
一句话总结

MTST 引入一个多分支、多分辨率的基于补丁的 transformer,采用相对位置编码,以建模长期的多变量时间序列的多样性模式,在基准数据集上实现了最先进的结果。

ABSTRACT

The performance of transformers for time-series forecasting has improved significantly. Recent architectures learn complex temporal patterns by segmenting a time-series into patches and using the patches as tokens. The patch size controls the ability of transformers to learn the temporal patterns at different frequencies: shorter patches are effective for learning localized, high-frequency patterns, whereas mining long-term seasonalities and trends requires longer patches. Inspired by this observation, we propose a novel framework, Multi-resolution Time-Series Transformer (MTST), which consists of a multi-branch architecture for simultaneous modeling of diverse temporal patterns at different resolutions. In contrast to many existing time-series transformers, we employ relative positional encoding, which is better suited for extracting periodic components at different scales. Extensive experiments on several real-world datasets demonstrate the effectiveness of MTST in comparison to state-of-the-art forecasting techniques.

研究动机与目标

  • 在长期预测中证明需要建模多尺度时间模式。
  • 提出一个使用不同补丁大小以捕捉多样频率的多分支 MTST。
  • 采用相对位置编码,以更好地捕捉周期性分量。
  • 在多个真实数据集上展示更优的预测性能,并提供消融研究以证明设计选择。

提出的方法

  • 构建 MTST,含 N 层,每层包含 B_n 个分支,用不同的补丁大小对输入进行标记。
  • 在每个分支中,使用带有相对位置编码的自注意力对补丁级标记进行处理。
  • 在每个 MTST 层中融合分支表示,形成下一层的共享嵌入。
  • 独立处理每个时间序列通道(通道独立性),并可扩展到跨通道相关性。
  • 训练以最小化 MSE,使用 Adam;对输入应用实例归一化,输出进行去归一化。

实验结果

研究问题

  • RQ1多分辨率、多分支的 transformer 是否相较单一分辨率的补丁模型在长期预测中具有提升?
  • RQ2在 MTST 中使用相对位置编码与绝对位置编码相比,其影响是什么?
  • RQ3对包含或排除高分辨率/低分辨率分支的消融对性能有何影响?
  • RQ4与最先进的基线相比,MTST 在多样化真实数据集和预测时域上的表现如何?

主要发现

DatasetTMTST_MSEMTST_MAEPatchTST_MSEPatchTST_MAEDLinear_MSEDLinear_MAEMICN_MSEMICN_MAETimesNet_MSETimesNet_MAEFedformer_MSEFedformer_MAEAutoformer_MSEAutoformer_MAEPyraformer_MSEPyraformer_MAE
Traffic960.3560.2440.3670.2510.4100.2820.4730.2930.5950.3180.5760.3590.5970.3712.0850.468
Traffic1920.3750.2510.3850.2590.4230.2870.4830.2980.6150.3260.6100.3800.6070.3820.8670.467
Traffic3360.3860.2560.3980.2650.4360.2960.4910.3030.6160.3260.6080.3750.6230.3870.8690.469
Traffic7200.4250.2790.4340.2870.4660.3150.5590.3270.6550.3530.6210.3750.6390.3950.8810.473
Electricity960.1270.2220.1300.2220.1400.2370.1570.2660.1780.2840.1860.3020.1960.3130.3860.449
Electricity1920.1440.2380.1480.2400.1530.2490.1750.2870.1870.2890.1970.3110.2110.3240.3860.443
Electricity3360.1620.2560.1670.2610.1690.2670.2000.3080.2080.3070.2130.3280.2140.3270.3780.443
Electricity7200.1990.2890.2020.2910.2030.3010.2280.3380.2450.3210.2330.3440.2360.3420.3760.445
Weather960.1500.1990.1520.1990.1760.2370.1780.2490.1630.2190.2380.3140.2490.3290.8960.556
Weather1920.1940.2400.1970.2430.2110.2690.2430.2690.2110.2590.2750.3290.3250.3700.6220.624
Weather3360.2460.2810.2490.2830.2650.3190.2780.3380.2860.3110.3390.3770.3510.3910.7390.753
Weather7200.3190.3330.3200.3350.3230.3620.3200.3600.3590.3630.3890.4090.4150.4261.0040.934
ETTh1960.3580.3900.3750.3990.3750.3990.4130.4420.4210.4400.3760.4150.4350.4460.6640.612
ETTh11920.3960.4140.4140.4210.4050.4160.4510.4620.5110.4980.4230.4460.4560.4570.7900.681
ETTh13360.3910.4200.4310.4360.4390.4430.5560.5280.4840.4780.4440.4620.4860.4870.8910.738
ETTh17200.4300.4570.4490.4660.4720.4900.6580.6070.5540.5270.4690.4920.5150.5170.9630.782
ETTh2960.2570.3260.2740.3360.2890.3530.3030.3640.3660.4170.3320.3740.3320.3680.6450.597
ETTh21920.3090.3610.3390.3790.3830.4180.4030.4460.4260.4470.4070.4460.4260.4340.7880.683
ETTh23360.3020.3660.3310.3800.4480.4650.6030.5500.4060.4350.4000.4470.4770.4790.9070.747
ETTh27200.3720.4160.3790.4220.6050.5511.1060.8520.4270.4570.4120.4690.4530.4900.9630.783
ETTm1960.2860.3380.2900.3420.2990.3430.3080.3600.3560.3850.3260.3900.5100.4920.5430.510
ETTm11920.3270.3660.3320.3690.3350.3650.3430.3840.4520.4280.3650.4150.5140.4950.5570.537
ETTm13360.3620.3890.3660.3920.3690.3860.3950.4110.4190.4250.3920.4250.5100.4920.7540.655
ETTm17200.4140.4210.4200.4240.4250.4210.4270.4340.4520.4510.4460.4580.5270.4930.9080.724
ETTm2960.1620.2510.1650.2550.1670.2600.1690.2680.1880.2760.1800.2710.2050.2930.4350.507
ETTm21920.2200.2910.2200.2920.2240.3030.2470.3330.2420.3100.2520.3180.2780.3360.7300.673
ETTm23360.2720.3260.2780.3290.2810.3420.2900.3510.3000.3460.3240.3640.3430.3791.2010.845
ETTm27200.3580.3790.3670.3850.3970.4210.4170.4340.3910.4030.4100.4200.4140.4193.6251.451
  • MTST 在 7 个数据集、4 个时域和 2 个指标上达到最优性能。
  • MTST 在 28 次 MSE 比较中有 27 次优于 PatchTST,且具有统计显著性。
  • 消融显示去除低分辨率或高分辨率分支都会降低性能,验证了多尺度建模的价值。
  • 相对位置编码始终比绝对编码提升预测准确性。
  • 回看窗口分析和定性可视化支持 MTST 在捕捉多尺度时间结构方面的优势。

更好的研究,从现在开始

从论文设计到论文写作,大幅缩短您的研究时间。

无需绑定信用卡

本解读由 AI 生成,并经人工编辑审核。