Graph-based Virtual Sensing from Sparse and Partial Multivariate Observations (ICLR 2024)

Authors: Giovanni De Felice, Andrea Cini, Daniele Zambon, Vladimir Gusev, Cesare Alippi

Code and official repository for the paper "Graph-based Virtual Sensing from Sparse and Partial Multivariate Observations". In this paper, we propose a graph-based methodology for tackling virtual sensing in a sparse and multivariate setting. The present code implementation is based on Torch Spatiotemporal, a library built to accelerate research on neural spatiotemporal data processing methods.

Guidelines for executing the code. (Tested with Python 3.10 on Ubuntu 22.04.3 LTS)

1. install dependencies:

conda update conda
conda env create -f conda_env_linux.yml  
conda activate ggnet

2. install torch spatiotemporal

open a terminal in the directory where the 'README.txt' is located

git clone https://github.com/TorchSpatiotemporal/tsl.git
cd tsl
pip install -e .

3. create datasets:

• climate: use the script data/NASA_data/build_dataset.py. This will download data from the API. Then save the output into data/NASA_data/clmDaily.pkl or data/NASA_data/clmHourly.pkl (may take some time to complete)

• photovoltaic: First, install install xarray: pip install xarray. Download coordinates.nc and module_00.tar.gz (~86 Gb) from https://scholarsphere.psu.edu/resources/dacba268-d084-4e0e-a674-670217c59891 and place both into the data/Photovoltaic folder. Finally, modify and use the script data/Photovoltaic/build_dataset.py. One dataset with 50 nodes is provided in the code.

• feel free to contact me for support or requesting built datasets

4. choose model and dataset in config.yaml

5. choose setting, e.g., number of epochs, in default.yaml

(optional) to enable logging, uncomment and personalize the wandb configurations into default.yaml

6. run the code: python run.py config=config