yorkjiang1998/gnn_wind_power_forecasting

In this project, we develop physics-induced graph neural networks for spatio-temporal, probabilistic wind power forecasting. The newly developed methods will be benchmarked against other state-of-the-art forecasting models.

Physics-Induced Graph Neural Network for Probabilistic Spatio-Temporal Forecasting of Wind Power

In this project, we develop physics-induced graph neural networks for spatio-temporal, probabilistic wind power forecasting. The newly developed methods will be benchmarked against other state-of-the-art forecasting models.

Getting Started

Installing Dependencies

Some dependencies require Microsoft Visual C++ 14.0, so make sure to have this installed on your machine first.

To then install the project's dependencies, create a new conda environment, activate it, and first run

conda install pip

Then, in the project folder root, run

pip install -r requirements.txt

Due to some conflicting package dependencis, we recommend to then run separately

pip install torch-geometric-temporal --no-deps

In PyTorch Geometric Temporal, there is currently a bug described in this issue, which must be fixed manually in the package file tsagcn.py.

Running a Model

To train a model, go to the project root and run

python run.py

You can optionally specify the parameters below. If not specified, the parameter's default will be used.

• --correlation_threshold: correlation threshold of when to connect wind sites with an edge in the input graph, options [05, 06, 07, 08, 09], default: 08.
• --data: dataset to use, options: [aemo, kelmarsh], default: kelmarsh.
• --device: torch device, options: [cpu, cuda], default: cpu.
• --model: model to use, options: [mlp, temporal_gnn, tgcn], default: mlp.
• --num_epochs: number of epochs, default: 10.
• --num_timesteps_in: length (number of consecutive data points) of the look back window, default: 12.
• --num_timesteps_in: number of consecutive data points to predict, default: 12.
• --train_data_amount: percentage of training data to use for training, options: integer between 1 and 100, default: 10.
• --use_wandb: whether to track experiment in Weights & Biases, see https://wandb.ai/site, default: False.
• --wandb_project_name: name of Weights & Biases project to initialize, default: sulphur-crested-cockatoo.

You can also run

python run.py --help

to see all valid parameter choices and its defaults.

Data

The datasets used here are

• 10min Supervisory Control and Data Acquisition (SCADA) data from the six wind turbines at the Kelmarsh wind farm [1], available under a Creative Commons Attribution 4.0 International license, and
• 5min SCADA data from 2022 to 2023 for 67 Australian wind farms, retrieved from the Nemweb archive from AEMO [2], which is free to use for any purpose, when accurately and appropriately attributing AEMO as the author, see the AEMO Copyright Permissions.

Acknowledgment

We appreciate the following GitHub repository for its publicly available code and methods, which we use as benchmark methods for our architecture:

https://github.com/LarsBentsen/FFTransformer

References

[1] Plumley, C. (2022). Kelmarsh wind farm data (0.1.0). Zenodo. https://doi.org/10.5281/zenodo.8252025

[2] AEMO (2024). Nemweb Archive Reports. National Electricity Market Web. https://nemweb.com.au/Reports/Archive/