Beyond Fixed Variables: Expanding-variate Time Series Forecasting via Flat Scheme and Spatial-temporal Focal Learning

:house: STEV

TL; DR

We introduce EVTSF, an emerging task that targets a frequently overlooked aspect of CPSs — their evolution through sensing expansion. To address this challenge, we propose STEV, a pioneering framework tailored to the unique demands of EVTSF.

Python library

pip install -r requirements.txt

Data Preparation

We sourced from three public multivariate time series datasets: Electricity,PeMS, and Weather. You need to download these datasets first, then run the data_process scripts to obtain the expanding-variate time series datasets.

For convenience, we also uploaded the processed dataset to Baidu Driver with the password "9432".

You can also generate the datasets using the scripts.

# cd data/script

# Generating EElectricity
python process_elc.py
python process_elc_oracle.py

# Generating EPeMS
python process_pems.py
python process_pems_spatial.py
python process_pems_random.py

# Generating EWeather
python process_weather.py
python process_weather_oracle.py

---

Run STEV

Training & Validating & Infering

# EElectricity
python main.py common=stev_electricity

# EPeMS
python main.py common=stev_pems

# EWeather
python main.py common=stev_weather

For all experiments

We also provide a script for running all EVTSF models including STEV, Uni methods, FPTM methods. Moreover, the Oracle experiment is included.

bash run.sh

Only for Infering with trained weights

python infer.py # modify the weight folder path in infer.py