AI4RecWind (Artificial Intelligence for the Reconstruction of Wind)

by Climatoc-Lab

AI4RecWind provides tools for reconstructing missing windspeed observations in AEMET’s historical dataset using a U-Net with partial convolutions (CRAI model).
This implementation builds on the CRAI software developed by DKRZ and adapts it to reconstruct daily windspeed maps in Spain.

---

Overview

• 🧠 Model architecture: Based on a partial convolution U-Net (PCNN) with optional LSTM, GRU, and Attention mechanisms.
• 🗺️ Use case: Filling gaps in daily windspeed maps from AEMET’s station network.

---

1. Environment Setup

You can set up the Python environment using Conda.

🧪 CPU Environment

conda env create -f environment.yml
conda activate crai

🚀 GPU Environment (CUDA)

conda env create -f environment-cuda.yml
conda activate crai

---

2. Installing CRAI

With the environment activated, install the climatereconstructionAI package:

pip install model/climatereconstructionAI

---

3. Reconstruction (Validation)

Once the model is installed and the environment is ready, you can reconstruct missing values using a pre-trained model (typically saved as best.pth):

bash run_eval_CRAI.sh

🛠️ Setup Instructions

Before running the script, ensure run_eval_CRAI.sh is updated with:

• Paths to:
  • Input data (input_data/)
  • Masks (defining valid and missing pixels)
  • Model checkpoints
  • Output directory (where infilled data will be saved)
• Configuration details such as:
  • File names of inputs and masks
  • Output file name
  • Device selection: cuda (for GPU) or cpu

📁 Input Files

• input_data/ should include:
  • test/: test data to reconstruct
  • masks/: observation masks (1 = valid, 0 = missing)
  • steady_mask_reversed: inverted land/sea mask (1 = sea, 0 = land), required for evaluation (provided in the repository for Spain)

⚙️ Configuration File

The script uses evaluation_spain.inp, which defines:

• The variable to reconstruct
• Model hyperparameters
• Number of partitions (to manage GPU memory usage: the smaller the VRAM of the gpu the larger the number of the partitions is needed)

📤 Outputs

After evaluation, results are saved in the evaluation/ folder:

• name_output.nc: Final reconstruction, merging model output with original observations
• name_infilled.nc: Raw model prediction (infilled data only)
• name_gt.nc, name_image.nc, name_mask.nc: Supporting files (see evaluation/Output of CRAI evaluation.txt)

---

4. Training

If you wish to train your own model from scratch:

📁 Input Structure

• Place training and validation data in input_data/train/ and input_data/val/
• Validation files must have the same names as training files
• Include:
  • Gridded data files (complete datasets)
  • Observation masks for each timestamp in input_data/masks/
    (reflecting which grid points are valid/missing at each time)
  • Steady land/sea mask: 1 = land, 0 = sea (provided in the repository for Spain)
  • Inverted land/sea mask (used in evaluation)

▶️ Launching Training

Navigate to the execution/ folder and run:

bash run_train_CRAI.sh

Edit this script to specify:

• Paths to input data and masks
• Device to use (cpu or cuda)
• Output directories for logs and checkpoints

📄 Training Configuration

Defined in ws_crai_training.inp, including:

• Batch size
• Learning rate
• Model architecture (layers, attention, etc.)
• Variable names and masks used

📦 Training Outputs

• Logs are saved in the logs/ folder
• Model checkpoints in snapshots/ckpt/, including:
  • best.pth: best-performing model based on validation loss

---

5. CRAI Model Origin

This software is based on the CRAI model developed by the Data Analysis Group led by Christopher Kadow at the Deutsches Klimarechenzentrum (DKRZ).

📚 Reference

Kadow et al. (2020), Nature Geoscience
DOI: 10.1038/s41561-020-0582-5

👥 Contributors

Maintained by the Climate Informatics and Technology Group at DKRZ

• Previous contributing authors:: Naoto Inoue, Christopher Kadow, Stephan Seitz
• Current contributing authors: Johannes Meuer, Maximilian Witte, Étienne Plésiat

🔑 License

CRAI is licensed under the terms of the BSD 3-Clause license.

---