notesum.ai

Published at November 3

PSformer: Parameter-efficient Transformer with Segment Attention for Time Series Forecasting

cs.LG

cs.AI

Released Date: November 3, 2024

Authors: Yanlong Wang¹, Jian Xu¹, Fei Ma², Shao-Lun Huang¹, Danny Dongning Sun³, Xiao-Ping Zhang¹

Aff.: ¹Shenzhen International Graduate School, Tsinghua University; ²Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ); ³Peng Cheng Laboratory

Arxiv: http://arxiv.org/abs/2411.01419v1

Dataset	Horizon	PSformer		SAMformer	TSMixer	ModernTCN	RLinear
Dataset	Horizon	Full	Encoder	SAMformer	TSMixer	ModernTCN	RLinear
ETTh1	H=96	52,416	3,168	50,272	124,142	876,064	49248
	H=192	101,664	3,168	99,520	173,390	1,662,592	98496
	H=336	175,536	3,168	173,392	247,262	2,842,384	172368
	H=720	372,528	3,168	369,904	444,254	5,988,496	369360
Relative Size (Avg)		1.0	0.014	0.987	1.408	16.192	0.948
Weather	H=96	58,752	9,504	50,272	121,908	2,709,280	49248
	H=192	108,000	9,504	99,520	171,156	3,495,808	98496
	H=336	181,872	9,504	173,392	245,028	4,675,600	172368
	H=720	378,864	9,504	369,904	442,020	7,821,712	369360
Relative Size (Avg)		1.0	0.039	0.953	1.347	25.708	0.948
Traffic	H=96	58,752	9,504	50,272	793,424	822,018,208	49248
	H=192	108,000	9,504	99,520	842,672	822,804,736	98496
	H=336	181,872	9,504	173,392	916,544	823,984,528	172368
	H=720	378,864	9,504	369,904	1,113,536	827,130,640	369360
Relative Size (Avg)		1.0	0.039	0.953	5.040	4530.574	0.948