A systematic synthesis of sky image enhancement techniques for ground-based solar irradiance forecasting
A systematic synthesis of sky image enhancement techniques for ground-based solar irradiance forecasting.
- Python 3.12.6
- Python packages defined in the pyproject.toml file
pip install -e .[dev]1. Sun mask
2. Irradiance channel (with Sun mask)
3. Optical flow
Optical flow data generated using DIS method, with ultrafast preset, to enhance the input image at time t as additional channels. (a) history image (at time t − 15); (b) optical flow in the X-axis generated between an input image and previous images (at t − 15, t − 10, t − 5); (c) optical flow in the Y-axis generated between an input image and previous images (at t − 15, t − 10, t − 5); (d) input image (at time t). Red and blue colours represent the direction of pixel displacement, respectively positive and negative, between the initial and the final frame in the x and y directions. Colour saturation stands for shift magnitude.
4. Cloud channel
Visualization of utilized cloud channel methods. For illustration purposes, the output masks were thresholded with hand-picked values to produce binary masks. (a) input image; (b) red-blue ratio (R2B); (c) red-blue difference (R-B); (d) normalized blue-red ratio (Norm. B/R).
The dataset used in this study is based on the Folsom dataset. To make the data preparation and follow-up steps easier to reproduce, these steps were described as a directed acyclic graph (DAG) using the DVC package. This pipeline is stored in a dvc.yaml file, while data are stored in the data directory.
For study reproducibility, Zenodo repository with data splits and pretrained models was created and is available at https://zenodo.org/record/15792283.
1. DVC DAG (Directed Acyclic Graph)
dvc dag +----------+
*| get_data |**
***** +----------+ ******
****** *****
***** *****
*** ******
+--------------------+ ***
| convert_timestamps |***** *
+--------------------+ *********** *
*** *** *********** *
*** *** *********** *
** ** ****** *
** ** +-----------------+
* * | clean_dataframe |
* * +-----------------+
* * *** ***
* * ** ***
* * ** **
*** *** +----------------+ ***
***** ***** | export_periods | *****
***** ****** ****+----------------+ *****
***** ****** * *****
***** ****** ***** * *****
*** *** *** * ***
+---------------------+ +------------------+
| export_eval_periods | | train_forecaster |
+---------------------+ +------------------+2. Data structure
data/
├── raw
├── prepared
└── eval- raw - raw irradiance CSV and sky images extracted directly from files downloaded from Folsom Zenodo page
- prepared - processed data (filtered, timestamp moved to local timezone, corrupted images removed, and grouped into periods)
- eval - small subset extracted for evaluation purposes
- Train
python -m scripts.train_forecaster --data-root data/prepared --periods-filename periods.pickle- Test
Change test_only: True and provide path to checkpoint restore_from_ckpt: path/to/checkpoint. Then call train script.
- Evaluate
Firstly export PyTorch checkpoint to ONNX format, for example using test procedure with export_to_onnx: True.
python -m scripts.evaluate \
--model-path /path/to/model \
--dataset-path data/eval/images \
--eval-periods-path data/eval/eval_periods.pickle \
--provider cpu \
--add-sun-mask \
--add-irradiance-channel \
--cloud-mask-method normalized_blue_red_ratio \
--optical-flow dis
