SCLformer

SCLformer for LSTF problem

Introduce

SCLformer is short for Self-adaptive Convolution Linear Transformer. It is a transformer-based model for LSTF problem. The model is based on the paper. The model is implemented by Pytorch.

Usage

Dataset

Dataset can be downloaded from here.

Result

The result of SCLformer is shown as follows:

Multivariable condition:

Methods		SCLformer		Informer		LSTM		ARMA		DeepAR
Metric		MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE
ETTh1	96	0.903	0.73	0.906	0.746	1.201	0.861	0.846	0.607	0.833	0.701
	192	1.024	0.797	0.843	0.694	1.23	0.831	0.861	0.62	0.939	0.749
	336	0.999	0.785	1.157	0.849	1.292	0.921	0.875	0.635	1.088	0.785
	720	1.065	0.835	1.231	0.887	1.126	0.839	0.884	0.653	1.078	0.846
ETTh2	96	1.767	1.097	2.723	1.404	2.193	1.179	3.154	1.353	2.283	1.196
	192	3.516	1.605	5.904	2.11	3.236	1.402	3.166	1.357	3.355	1.565
	336	3.56	1.586	3.511	1.553	2.533	1.274	3.149	1.351	2.681	1.405
	720	4.232	1.762	3.283	1.56	3.498	1.552	3.113	1.342	3.049	1.333
ETTm1	96	0.655	0.585	0.556	0.529	1.124	0.8222	0.865	0.619	0.779	0.701
	192	0.697	0.631	0.872	0.69	1.261	0.898	0.871	0.621	0.805	0.712
	336	0.865	0.715	0.933	0.739	1.14	0.843	0.882	0.631	0.83	0.724
	720	0.927	0.751	0.918	0.732	1.184	0.869	0.899	0.644	0.889	0.75
ETTm2	96	0.696	0.607	0.539	0.522	1.134	0.887	3.121	1.343	1.026	0.816
	192	1.083	0.753	1.05	0.769	2.366	1.299	3.128	1.346	1.498	0.952
	336	0.885	0.726	1.451	0.935	2.612	1.321	3.142	1.349	1.997	1.136
	720	1.835	1.042	2.441	1.172	3.127	1.508	3.152	1.352	2.368	1.212
Weather	96	0.755	0.665	0.488	0.497	0.554	0.548	0.618	0.557	0.487	0.5
	192	0.795	0.687	0.581	0.552	0.604	0.581	0.642	0.579	0.528	0.532
	336	0.676	0.613	0.604	0.571	0.617	0.589	0.647	0.588	0.546	0.547
	720	0.746	0.648	0.614	0.583	0.652	0.61	0.672	0.61	0.638	0.603
ECL	96	0.336	0.404	0.303	0.395	1.013	0.837	0.584	0.572	0.53	0.541
	192	0.311	0.39	0.291	0.378	0.932	0.793	0.582	0.574	0.492	0.522
	336	0.306	0.387	0.302	0.392	0.962	0.801	0.586	0.582	0.488	0.522
	720	0.314	0.395	0.381	0.447	0.953	0.804	0.603	0.599	0.505	0.514

Singlevariable condition:

Methods		SCLformer		Informer		LSTM		ARMA		DeepAR
Metric		MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE
ETTh1	96	0.903	0.73	0.906	0.746	1.201	0.861	0.846	0.607	0.833	0.701
	192	1.024	0.797	0.843	0.694	1.23	0.831	0.861	0.62	0.939	0.749
	336	0.999	0.785	1.157	0.849	1.292	0.921	0.875	0.635	1.088	0.785
	720	1.065	0.835	1.231	0.887	1.126	0.839	0.884	0.653	1.078	0.846
ETTh2	96	1.767	1.097	2.723	1.404	2.193	1.179	3.154	1.353	2.283	1.196
	192	3.516	1.605	5.904	2.11	3.236	1.402	3.166	1.357	3.355	1.565
	336	3.56	1.586	3.511	1.553	2.533	1.274	3.149	1.351	2.681	1.405
	720	4.232	1.762	3.283	1.56	3.498	1.552	3.113	1.342	3.049	1.333
ETTm1	96	0.655	0.585	0.556	0.529	1.124	0.8222	0.865	0.619	0.779	0.701
	192	0.697	0.631	0.872	0.69	1.261	0.898	0.871	0.621	0.805	0.712
	336	0.865	0.715	0.933	0.739	1.14	0.843	0.882	0.631	0.83	0.724
	720	0.927	0.751	0.918	0.732	1.184	0.869	0.899	0.644	0.889	0.75
ETTm2	96	0.696	0.607	0.539	0.522	1.134	0.887	3.121	1.343	1.026	0.816
	192	1.083	0.753	1.05	0.769	2.366	1.299	3.128	1.346	1.498	0.952
	336	0.885	0.726	1.451	0.935	2.612	1.321	3.142	1.349	1.997	1.136
	720	1.835	1.042	2.441	1.172	3.127	1.508	3.152	1.352	2.368	1.212
Weather	96	0.755	0.665	0.488	0.497	0.554	0.548	0.618	0.557	0.487	0.5
	192	0.795	0.687	0.581	0.552	0.604	0.581	0.642	0.579	0.528	0.532
	336	0.676	0.613	0.604	0.571	0.617	0.589	0.647	0.588	0.546	0.547
	720	0.746	0.648	0.614	0.583	0.652	0.61	0.672	0.61	0.638	0.603
ECL	96	0.336	0.404	0.303	0.395	1.013	0.837	0.584	0.572	0.53	0.541
	192	0.311	0.39	0.291	0.378	0.932	0.793	0.582	0.574	0.492	0.522
	336	0.306	0.387	0.302	0.392	0.962	0.801	0.586	0.582	0.488	0.522
	720	0.314	0.395	0.381	0.447	0.953	0.804	0.603	0.599	0.505	0.514

More info are shown in the paper