Recognizing Dominant Patterns for Long-term Time Series Forecasting

Datasets

This repository only contains the code of PRNet. Seven datasets are available here, including Electricity, ETT, Exchange, QPS, Solar, Traffic and Weather. They should be firstly unzipped and moved into the dataset folder.

Dataset Lists:

• Electricity
  • LD2011_2014.csv
  • LD2011_2014_h.csv
• ETT
  • ETTh.csv
  • ETTm.csv
• Exchange
  • exchange_rate.csv
• QPS
  • HQPS.csv
  • MQPS.csv
• Solar
  • solar_Alabama.csv
  • solar_Alabama_h.csv
• Traffic
  • PeMS.csv
• Weather
  • mpi_roof.csv
  • mpi_roof_h.csv

Run

You can run main.py to reproduce the experiment. Below is an example of running the Traffic dataset with pred_len = 24.

python3 main.py -cuda_id 0 -dataset Traffic -pred_len 24

There are seven dataset names:

Electricity ETT Exchange QPS Solar Traffic Weather

and their hyperparameters are listed in files/configs.json.

The data_loader loads the datasets, model contains the code of PRNet, and files/networks saves the cases trained by ourselves. Below is the experiment result.

We have also unified the sampling interval of datasets except for Exchange to 1 hour for better comparison. You can set the parameter interval to 'H' to evaluate these unified datasets.

Visualization

There is also a visualize function, which is used for visualizing the forecasting result. The package files/figures saves eight forecasting results, you can also run the trained models by yourself.

Forecasting Accuracy

Dataset	L	L		2L		3.5L		7.5L
		MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE
Electricity	96	0.1333	0.2263	0.1572	0.2453	0.1604	0.2453	0.1638	0.2513
ETT	96	0.2052	0.2988	0.2437	0.3291	0.2813	0.3571	0.3366	0.3987
Exchange	30	0.0204	0.0961	0.0371	0.1326	0.0630	0.1765	0.1357	0.2635
QPS	60	0.0279	0.0928	0.0583	0.1474	0.1294	0.2297	0.3233	0.3853
Solar	288	0.1836	0.2421	0.1980	0.2558	0.2066	0.2594	0.2069	0.2557
Traffic	24	0.3239	0.3007	0.3541	0.3226	0.3781	0.3380	0.3794	0.3365
Weather	144	0.3405	0.3328	0.4138	0.3923	0.4755	0.4386	0.5455	0.4807
ETTh	24	0.2134	0.3046	0.2458	0.3301	0.2834	0.3576	0.3328	0.3969

Contact

If you have any questions or suggestions for our paper or codes, please contact us. Email: hanwen_hu@sjtu.edu.cn.