SCINet

This is the original PyTorch implementation of the following work: Time Series is a Special Sequence: Forecasting with Sample Convolution and Interaction. If you find this repository useful for your work, please consider citing it as follows:

@article{liu2021SCINet,
  title={Time Series is a Special Sequence: Forecasting with Sample Convolution and Interaction},
  author={Liu, Minhao and Zeng, Ailing and Lai, Qiuxia and Xu, Qiang},
  journal={arXiv preprint arXiv:2106.09305},
  year={2021}
}

Updates

[2021-09-17] SCINet v1.0 is released!

Features

• [x] Support 11 popular time-series forecasting datasets.
• [x] Provide all training logs.

To-do items

• Integrate GNN-based spatial models into SCINet for better performance on spatial-temporal time series. Our preliminary results show that this feature could result in considerable gains on the prediction accuracy of some datasets (e.g., PEMSxx).
• Generate probalistic forecasting results.

Stay tuned!

Dataset

We conduct the experiments on 11 popular time-series datasets, namely Electricity Transformer Temperature (ETTh1, ETTh2 and ETTm1) , Traffic, Solar-Energy, Electricity and Exchange Rate and PeMS (PEMS03, PEMS04, PEMS07 and PEMS08), ranging from power, energy, finance and traffic domains.

Overall information of the 11 datasets

Datasets	Variants	Timesteps	Granularity	Start time	Task Type
ETTh1	7	17,420	1hour	7/1/2016	Multi-step
ETTh2	7	17,420	1hour	7/1/2016	Multi-step
ETTm1	7	69,680	15min	7/1/2016	Multi-step
PEMS03	358	26,209	5min	5/1/2012	Multi-step
PEMS04	307	16,992	5min	7/1/2017	Multi-step
PEMS07	883	28,224	5min	5/1/2017	Multi-step
PEMS08	170	17,856	5min	3/1/2012	Multi-step
Traffic	862	17,544	1hour	1/1/2015	Single-step
Solar-Energy	137	52,560	1hour	1/1/2006	Single-step
Electricity	321	26,304	1hour	1/1/2012	Single-step
Exchange-Rate	8	7,588	1hour	1/1/1990	Single-step

Get started

Requirements

Install the required package first:

cd SCINet
conda create -n scinet python=3.8
conda activate scinet
pip install -r requirements.txt

Dataset preparation

All datasets can be downloaded here. To prepare all dataset at one time, you can just run:

source prepare_data.sh

The data directory structure is shown as follows.

|-- datasets/
`-- |-- ETT-data/
    `-- |-- ETTh1.csv
        |-- ETTh2.csv
        |-- ETThm1.csv
`-- |-- financial/
    `-- |-- electricity.txt
        |-- exchange_rate.txt
        |-- solar_AL.txt
        |-- traffic.txt
`-- |-- PEMS/
    `-- |-- PEMS03.npz
        |-- PEMS04.npz
        |-- PEMS07.npz
        |-- PEMS08.npz

Run training code

To facilitate reproduction, we provide the logs on the above datasets here in details. You can check the hyperparameters, training loss and test results for each epoch in these logs as well.

We follow the same settings of StemGNN for PEMS 03, 04, 07, 08 datasets, MTGNN for Solar, electricity, traffic, financial datasets, Informer for ETTH1, ETTH2, ETTM1 datasets. The detailed training commands are given as follows.

For PEMS dataset (All datasets follow Input 12, Output 12):

pems03

python run_pems.py --dataset PEMS03 --hidden-size 0.0625 --dropout 0.25 --model_name pems03_h0.0625_dp0.25

pems04

python run_pems.py --dataset PEMS04 --hidden-size 0.0625 --dropout 0 --model_name pems04_h0.0625_dp0

pems07

python run_pems.py --dataset PEMS07 --hidden-size 0.03125 --dropout 0.25 --model_name pems07_h0.03125_dp0.25

pems08

python run_pems.py --dataset PEMS08 --hidden-size 1 --dropout 0.5 --model_name pems08_h1_dp0.5

PEMS Parameter highlights

Parameter Name	Description	Parameter in paper	Default
dataset	Name of dataset	N/A	PEMS08
horizon	Horizon	Horizon	12
window_size	Look-back window	Look-back window	12
hidden-size	hidden expansion	h	1
layers	SCINet block layers	L	2
stacks	The number of SCINet block	K	1

For Solar dataset:

predict 3

python run_financial.py --dataset_name solar_AL --window_size 160 --horizon 3 --hidden-size 2  --lastWeight 0.5 --stacks 1 --layers 4 --lradj 2 --lr 1e-4 --dropout 0.25 --batch_size 1024 --model_name so_I160_o3_lr1e-4_bs1024_dp0.25_h2_s1l4_w0.5

predict 6

python run_financial.py --dataset_name solar_AL --window_size 160 --horizon 6 --hidden-size 2 --lastWeight 0.5 --stacks 2 --layers 4 --lradj 2 --lr 1e-4 --dropout 0.25 --batch_size 1024 --model_name so_I160_o6_lr1e-4_bs1024_dp0.25_h2_s2l4_w0.5 

predict 12

python run_financial.py --dataset_name solar_AL --window_size 160 --horizon 12 --hidden-size 2 --lastWeight 0.5 --stacks 2 --layers 4 --lradj 2 --lr 1e-4 --dropout 0.25 --batch_size 1024 --model_name so_I160_o12_lr1e-4_bs1024_dp0.25_h2_s2l4_w0.5

predict 24

python run_financial.py --dataset_name solar_AL --window_size 160 --horizon 24 --hidden-size 2 --lastWeight 0.5 --stacks 1 --layers 4 --lradj 2 --lr 1e-4 --dropout 0.25 --batch_size 1024 --model_name so_I160_o24_lr1e-4_bs1024_dp0.25_h2_s1l4_w0.5

For Electricity dataset:

predict 3

python run_financial.py --dataset_name electricity --window_size 168 --horizon 3 --hidden-size 8 --single_step 1 --stacks 2 --layers 3 --lr 9e-3 --dropout 0 --batch_size 32 --model_name ele_I168_o3_lr9e-3_bs32_dp0_h8_s2l3_w0.5 --groups 321

predict 6

python run_financial.py --dataset_name electricity --window_size 168 --horizon 6 --hidden-size 8 --single_step 1 --stacks 2 --layers 3 --lr 9e-3 --dropout 0 --batch_size 32 --model_name ele_I168_o6_lr9e-3_bs32_dp0_h8_s2l3_w0.5 --groups 321

predict 12

python run_financial.py --dataset_name electricity --window_size 168 --horizon 12 --hidden-size 8 --single_step 1 --stacks 2 --layers 3 --lr 9e-3 --dropout 0 --batch_size 32 --model_name ele_I168_o12_lr9e-3_bs32_dp0_h8_s2l3_w0.5 --groups 321

predict 24

python run_financial.py --dataset_name electricity --window_size 168 --horizon 24 --hidden-size 8 --single_step 1 --stacks 2 --layers 3 --lr 9e-3 --dropout 0 --batch_size 32 --model_name ele_I168_o24_lr9e-3_bs32_dp0_h8_s2l3_w0.5 --groups 321

For Traffic dataset (warning: 20,000MiB+ memory usage!):

predict 3

python run_financial.py --dataset_name traffic --window_size 168 --horizon 3 --hidden-size 2 --single_step 1 --stacks 2 --layers 3 --lr 5e-4 --dropout 0.25 --batch_size 16 --model_name traf_I168_o3_lr5e-4_bs16_dp0.25_h2_s2l3_w1.0

predict 6

python run_financial.py --dataset_name traffic --window_size 168 --horizon 6 --hidden-size 2 --single_step 1 --stacks 2 --layers 2 --lr 5e-4 --dropout 0.25 --batch_size 16 --model_name traf_I168_o6_lr5e-4_bs16_dp0.25_h2_s2l2_w1.0

predict 12

python run_financial.py --dataset_name traffic --window_size 168 --horizon 12 --hidden-size 1 --single_step 1 --stacks 2 --layers 3 --lr 5e-4 --dropout 0.25 --batch_size 16 --model_name traf_I168_o12_lr5e-4_bs16_dp0.25_h1_s2l3_w1.0

predict 24

python run_financial.py --dataset_name traffic --window_size 168 --horizon 24 --hidden-size 2 --single_step 1 --stacks 2 --layers 2 --lr 5e-4 --dropout 0.5 --batch_size 16 --model_name traf_I168_o24_lr5e-4_bs16_dp0.5_h2_s2l2_w1.0

For Exchange rate dataset:

predict 3

python run_financial.py --dataset_name exchange_rate --window_size 168 --horizon 3 --hidden-size 0.125 --lastWeight 0.5 --stacks 1 --layers 3 --lr 5e-3 --dropout 0.5 --batch_size 4 --model_name ex_I168_o3_lr5e-3_bs4_dp0.5_h0.125_s1l3_w0.5 --epochs 150

predict 6

python run_financial.py --dataset_name exchange_rate --window_size 168 --horizon 6 --hidden-size 0.125 --lastWeight 0.5 --stacks 1 --layers 3 --lr 5e-3 --dropout 0.5 --batch_size 4 --model_name ex_I168_o6_lr5e-3_bs4_dp0.5_h0.125_s1l3_w0.5 --epochs 150

predict 12

python run_financial.py --dataset_name exchange_rate --window_size 168 --horizon 12 --hidden-size 0.125 --lastWeight 0.5 --stacks 1 --layers 3 --lr 5e-3 --dropout 0.5 --batch_size 4 --model_name ex_I168_o12_lr5e-3_bs4_dp0.5_h0.125_s1l3_w0.5 --epochs 150

predict 24

python run_financial.py --dataset_name exchange_rate --window_size 168 --horizon 24 --hidden-size 0.125 --lastWeight 0.5 --stacks 1 --layers 3 --lr 5e-3 --dropout 0.5 --batch_size 4 --model_name ex_I168_o24_lr5e-3_bs4_dp0.5_h0.125_s1l3_w0.5 --epochs 150

Financial Parameter highlights

Parameter Name	Description	Parameter in paper	Default
dataset_name	Data name	N/A	exchange_rate
horizon	Horizon	Horizon	3
window_size	Look-back window	Look-back window	168
batch_size	Batch size	batch size	8
lr	Learning rate	learning rate	5e-3
hidden-size	hidden expansion	h	1
layers	SCINet block layers	L	3
stacks	The number of SCINet block	K	1
lastweight	Loss weight of the last frame	Loss weight ($\lambda$)	1.0

For ETTH1 dataset:

multivariate, out 24

python run_ETTh.py --data ETTh1 --features M  --seq_len 48 --label_len 24 --pred_len 24 --hidden-size 4 --stacks 1 --layers 3 --lr 5e-3 --batch_size 16 --dropout 0.5 --model_name etth1_M_I48_O24_lr5e-3_bs16_dp0.5_h4_s1l3

multivariate, out 48

python run_ETTh.py --data ETTh1 --features M  --seq_len 96 --label_len 48 --pred_len 48 --hidden-size 4 --stacks 1 --layers 3 --lr 0.009 --batch_size 16 --dropout 0.25 --model_name etth1_M_I96_O48_lr0.009_bs16_dp0.25_h4_s1l3

multivariate, out 168

python run_ETTh.py --data ETTh1 --features M  --seq_len 336 --label_len 168 --pred_len 168 --hidden-size 4 --stacks 1 --layers 3 --lr 5e-4 --batch_size 32 --dropout 0.5 --model_name etth1_M_I336_O168_lr5e-4_bs32_dp0.5_h4_s1l3

multivariate, out 336

python run_ETTh.py --data ETTh1 --features M  --seq_len 336 --label_len 336 --pred_len 336 --hidden-size 1 --stacks 1 --layers 4 --lr 1e-4 --batch_size 512 --dropout 0.5 --model_name etth1_M_I336_O336_lr1e-4_bs512_dp0.5_h1_s1l4

multivariate, out 720

python run_ETTh.py --data ETTh1 --features M  --seq_len 736 --label_len 720 --pred_len 720 --hidden-size 1 --stacks 1 --layers 5 --lr 5e-5 --batch_size 256 --dropout 0.5 --model_name etth1_M_I736_O720_lr5e-5_bs256_dp0.5_h1_s1l5

Univariate, out 24

python run_ETTh.py --data ETTh1 --features S  --seq_len 64 --label_len 24 --pred_len 24 --hidden-size 8 --stacks 1 --layers 3 --lr 0.007 --batch_size 64 --dropout 0.25 --model_name etth1_S_I64_O24_lr0.007_bs64_dp0.25_h8_s1l3

Univariate, out 48

python run_ETTh.py --data ETTh1 --features S  --seq_len 720 --label_len 48 --pred_len 48 --hidden-size 4 --stacks 1 --layers 4 --lr 0.0001 --batch_size 8 --dropout 0.5 --model_name etth1_S_I720_O48_lr0.0001_bs8_dp0.5_h4_s1l4

Univariate, out 168

python run_ETTh.py --data ETTh1 --features S  --seq_len 720 --label_len 168 --pred_len 168 --hidden-size 4 --stacks 1 --layers 4 --lr 5e-5 --batch_size 8 --dropout 0.5 --model_name etth1_S_I720_O168_lr5e-5_bs8_dp0.5_h4_s1l4

Univariate, out 336

python run_ETTh.py --data ETTh1 --features S  --seq_len 720 --label_len 336 --pred_len 336 --hidden-size 1 --stacks 1 --layers 4 --lr 1e-3 --batch_size 128 --dropout 0.5 --model_name etth1_S_I720_O336_lr1e-3_bs128_dp0.5_h1_s1l4

Univariate, out 720 ``` python run_ETTh.py --data ETT