Skip to content

Latest commit

 

History

History
99 lines (66 loc) · 5.81 KB

README.md

File metadata and controls

99 lines (66 loc) · 5.81 KB
Raw

Wind Farm GNNs

This repository is dedicated to the development of Graph Neural Networks (GNN) for predicting wind farm-wide power, local flow variables and loads. Our models aim to predict 10min average power, rotor averaged wind speed and effective turbulence intensity, along with damage equivalent loads (DELs) in blade flap- and edgewise directions, tower top tortional and bottom fore-aft and side-to-side. The use of GNNs allows for a layout-agnostic multivariate model. Code associated with the following papers:

  • "Flexible multi-fidelity framework for load estimation of wind farms through graph neural networks and transfer learning", paper here
  • "Local flow and loads estimation on wake-affected wind turbines using graph neural networks and PyWake", paper here
  • "Multivariate prediction on wake-affected wind turbines using graph neural networks", paper here

Features

  • Robust & parallelized data generation: we use Sobol sampling to generate robust datasets with realistic inflow conditions and randomized layouts. We also parallelize PyWake to quickly generate training data.
  • Multivariate GNN models: outputs of both power and loads. Three default message-passing implementations are provided.
  • Layout-agnostic: capable of working with any wind farm layout.
  • Fast: runs ~10x quicker than PyWake, makes it suitable for RL envs.
  • LoRA-based finetuning: have some higher fidelity simulations or different turbines? Use the provided finetuning framework to transfer to higher fidelities or different turbines with a minimal amount of data.

Getting Started

To get started with this project, clone the repository and install the required dependencies.

Prerequisites

Ensure you have Python 3.x installed on your system. You can download it from python.org or using Anaconda (Miniconda is the best for a barebones lightweight install).

Installation

  1. Clone the repository: git clone https://github.com/GDuthe/windfarm-gnn.git

  2. Install dependencies: pip install -r requirements.txt

Project structure

windfarm-gnn/
│
├── gnn_framework/
│ ├── train.py              # script to train the GNN model
│ ├── config.yml            # template training config file
│ ├── predict.py            # evaluation of the GNN model
│ ├── finetuning/           # fine-tuning submodule
│ │    ├── finetune.py      # fine-tuning script
│ │    ├── ft_config.yml    # fine-tuning config file
│ │ 
│ ├── data/                 # data submodule (dataset class & more)
│ ...
│
├── graphfarms/
│ ├── generate_graphs.py    # script to generate wind farm graphs
│ ├── config.yml            # template data generation config file
│ ...
│
├── notebooks/              # example usage notebooks
│ ...
│
├── plotting/               # plotting functions
│ ...

Usage

Generating wind farm graph data

Navigate to the graph_farms folder.

To generate some training graphs with random layouts using PyWake and realistic random inflow conditions, you must first set up a config yaml file, please take a look the provided template. Then, you can launch the graph generation script. For example you can use:

python3 generate_graphs.py -c custom_config.yml -nl 200 -ni 15 -d ./your/custom/dataset/path -t 6

This command will generate a dataset with 200 layouts, each being simulated with 15 different inflows, using 6 threads. Some other arguments can be set and are detailed in generate_graphs.py.

Training a Model

Navigate into the gnn_framework directory.

To train a model, you must first set up a config yaml file. Please use the provided config.yml template as a starting point. Then, use the following command to launch the training script:

python3 train.py -c your_config.yml

Finetuning a Model

If you would like to finetune your trained model on different data (higher fidelity, new turbines, etc.) or a specific layout, you can use the provided finetuning submodule. Three different finetuning methods are implemented (LoRA, all-layers, decoder only). You can download one of our pretrained models (a 4 layer GEN GNN) here.

Again you must first set up a config file, have a look at the template provided in finetuning/ft_config.yml. To launch the fintuning process, stay in the gnn_framework directory and run:

python3 -m finetuning.finetune -y your_ft_config.yml

Testing a Model

We provide a function to gather prediction results for a given test dataset in predict.py.

Citing

If you are using the code, please consider citing:

Duthé G, de N Santos F, Abdallah I, Weijtjens W, Devriendt C, Chatzi E. Flexible multi-fidelity framework for load estimation of wind farms through graph neural networks and transfer learning. Data-Centric Engineering. 2024;5:e29. doi:10.1017/dce.2024.35

Duthé, G., de Nolasco Santos, F., Abdallah, I., Réthore, P.É., Weijtjens, W., Chatzi, E. and Devriendt, C., 2023, May. Local flow and loads estimation on wake-affected wind turbines using graph neural networks and PyWake. In Journal of Physics: Conference Series (Vol. 2505, No. 1, p. 012014). IOP Publishing.

de Nolasco Santos, F., Duthé, G., Abdallah, I., Réthoré, P. É., Weijtjens, W., Chatzi, E., & Devriendt, C. (2024). Multivariate prediction on wake-affected wind turbines using graph neural networks. In EURODYN 2023. Proceedings of the XII International Conference on Structural Dynamics. IOP Publishing.