3 个月前

xPatch:基于指数季节性-趋势分解的双流时间序列预测

xPatch:基于指数季节性-趋势分解的双流时间序列预测

摘要

近年来,基于Transformer的模型在时间序列预测领域受到广泛关注。尽管这类模型通常展现出令人瞩目的预测效果,但由于其注意力机制的局限性,难以充分挖掘时间序列数据中的时序依赖关系。为此,本文提出一种新型双流架构——指数块(eXponential Patch,简称xPatch),该架构引入了指数分解机制。受经典指数平滑方法的启发,xPatch创新性地设计了季节-趋势指数分解模块。此外,本文还提出一种双流结构,包含基于MLP的线性流与基于CNN的非线性流,旨在探索在非Transformer架构中采用分块(patching)与通道独立性(channel-independence)技术所带来的优势。最后,我们设计了一种鲁棒的反正切损失函数(arctangent loss function)以及一种Sigmoid形式的学习率自适应调整策略,有效抑制过拟合现象,显著提升预测性能。相关代码已开源,可访问以下仓库获取:https://github.com/stitsyuk/xPatch。

代码仓库

stitsyuk/xpatch
官方
pytorch
GitHub 中提及

基准测试

基准方法指标
time-series-forecasting-on-electricity-192xPatch
MAE: 0.232
MSE: 0.140
time-series-forecasting-on-electricity-336xPatch
MAE: 0.249
MSE: 0.156
time-series-forecasting-on-electricity-720xPatch
MAE: 0.281
MSE: 0.190
time-series-forecasting-on-electricity-96xPatch
MAE: 0.217
MSE: 0.126
time-series-forecasting-on-etth1-192-1xPatch
MAE: 0.395
MSE: 0.376
time-series-forecasting-on-etth1-336-1xPatch
MAE: 0.415
MSE: 0.391
time-series-forecasting-on-etth1-720-1xPatch
MAE: 0.459
MSE: 0.442
time-series-forecasting-on-etth1-96-1xPatch
MAE: 0.379
MSE: 0.354
time-series-forecasting-on-etth2-192-1xPatch
MAE: 0.330
MSE: 0.275
time-series-forecasting-on-etth2-336-1xPatch
MAE: 0.360
MSE: 0.312
time-series-forecasting-on-etth2-720-1xPatch
MAE: 0.418
MSE: 0.384
time-series-forecasting-on-etth2-96-1xPatch
MAE: 0.297
MSE: 0.226
time-series-forecasting-on-ettm1-192-1xPatch
Accuracy: 0.355
MSE: 0.315
time-series-forecasting-on-ettm1-336-1xPatch
MAE: 0.376
MSE: 0.355
time-series-forecasting-on-ettm1-720-1xPatch
MAE: 0.411
MSE: 0.419
time-series-forecasting-on-ettm1-96-1xPatch
MAE: 0.330
MSE: 0.275
time-series-forecasting-on-ettm2-192-1xPatch
MAE: 0.280
MSE: 0.213
time-series-forecasting-on-ettm2-336-1xPatch
MAE: 0.315
MSE: 0.264
time-series-forecasting-on-ettm2-720-1xPatch
MAE: 0.363
MSE: 0.338
time-series-forecasting-on-ettm2-96-1xPatch
MAE: 0.240
MSE: 0.153
time-series-forecasting-on-exchange-192-1xPatch
MAE: 0.298
MSE: 0.178
time-series-forecasting-on-exchange-336-1xPatch
MAE: 0.418
MSE: 0.339
time-series-forecasting-on-exchange-720-1xPatch
MAE: 0.701
MSE: 0.867
time-series-forecasting-on-exchange-96-1xPatch
MAE: 0.197
MSE: 0.081
time-series-forecasting-on-illness-24-1xPatch
MAE: 0.638
MSE: 1.188
time-series-forecasting-on-illness-36-1xPatch
MAE: 0.653
MSE: 1.226
time-series-forecasting-on-illness-48-1xPatch
MAE: 0.686
MSE: 1.254
time-series-forecasting-on-illness-60-1xPatch
Accuracy: 0.773
MSE: 1.455
time-series-forecasting-on-solar-192-1xPatch
MAE: 0.216
MSE: 0.193
time-series-forecasting-on-solar-336-1xPatch
MAE: 0.224
MSE: 0.196
time-series-forecasting-on-solar-720-1xPatch
MAE: 0.219
MSE: 0.212
time-series-forecasting-on-solar-96-1xPatch
MAE: 0.197
MSE: 0.173
time-series-forecasting-on-traffic-192xPatch
MAE: 0.241
MSE: 0.377
time-series-forecasting-on-traffic-336xPatch
MAE: 0.243
MSE: 0.388
time-series-forecasting-on-traffic-720xPatch
MAE: 0.273
MSE: 0.437
time-series-forecasting-on-traffic-96xPatch
MAE: 0.233
MSE: 0.364
time-series-forecasting-on-weather-192xPatch
MAE: 0.227
MSE: 0.189
time-series-forecasting-on-weather-336xPatch
MAE: 0.260
MSE: 0.218
time-series-forecasting-on-weather-720xPatch
MAE: 0.315
MSE: 0.291
time-series-forecasting-on-weather-96xPatch
MAE: 0.185
MSE: 0.146

用 AI 构建 AI

从想法到上线——通过免费 AI 协同编程、开箱即用的环境和市场最优价格的 GPU 加速您的 AI 开发

AI 协同编程
即用型 GPU
最优价格
立即开始

Hyper Newsletters

订阅我们的最新资讯
我们会在北京时间 每周一的上午九点 向您的邮箱投递本周内的最新更新
邮件发送服务由 MailChimp 提供
xPatch:基于指数季节性-趋势分解的双流时间序列预测 | 论文 | HyperAI超神经