Super-resolution of wind data using machine learning techniques. The repository contains an implementation of two methods:
- Local Regression: Implementation of the algorithm in "Fast and Accurate Image Upscaling with Super-Resolution Forests" using a ridge regression and a random forest model.
- Denoising Diffusion Models: Implementation of the algorithm in "Denoising Diffusion Probabilistic Models" with modification to handle super-resolution from "Image Super-Resolution via Iterative Refinement". Trained model weights are available in the repository.
The functions unet.py and sr3.py are adapted from the implementation at: https://github.com/TeaPearce/Conditional_Diffusion_MNIST.
The functions energy_spectrum.py, kinetic_energy_initial.py, and kinetic_energy_initial_with_post_normalization.py are adapted from the implementation at: https://github.com/RupaKurinchiVendhan/WiSoSuper/blob/main/energy.py
The SR3 algorithm was originally put forth by Saharia et al.:
[1] C. Saharia, J. Ho, W. Chan, T. Salimans, D. J. Fleet, and M. Norouzi, “Image Super-Resolution Via Iterative Refinement,” IEEE Trans. Pattern Anal. Mach. Intell., pp. 1–14, 2022, doi: 10.1109/TPAMI.2022.3204461.
First, construct a python environment and set up the directory using the "environment.yml" file. If the environment.yml file doesn't work, you need at least the following packages from various sources:
- numpy
- scipy
- matplotlib
- scikit-learn
- scikit-image
- pytorch
- gdown
- tqdm
- p7zip
Next, run the bash script ''dataset.sh.'' This will download the NREL dataset as used by the WiSoSuper Resolution paper. NOTE: there is a lot of data here. If only intending to run testing and validation, the training dataset can be commented out and is unnecessary.
Models:
- llr.py: implementation of the local regression models: ridge regression and random forest.
- unet.py: implementation of the UNet architecture used for the SR3 regression and diffusion models.
- sr3.py: main implementation of the SR3 regression and diffusion models.
- diffusion_noise_conditioned.py: alternative implementation of the diffusion model, conditioned on noise rather than timesteps.
- Trained models are available in the ''models' folder.
Functions:
- util.py: contains useful functions for the other functions, including bicubic interpolation.
- exp.py: functions to carry out main experiments in the code.
- metrics.py: stores the implementations of most metrics.
- visualization.py: contains functions to plot example data samples or results.
- energy_spectrum.py: implementation of the kinetic energy spectrum analysis, used to create Fig. 3 in our report.
- energy_spectrum_postprocess.py: used to post-process the energy_spectrum.py saved data if wanting to plot after the analysis.
- cosine_similarity.py: implementation of the cosine similarity metrics and analysis, used to create Fig. 4 in our report.
