Our model performance for next-day anti-psoriasis-weighted ultraviolet forecasting with different input sizes (7, 14, or 21 days) on 2018 or 2019 ultraviolet radiation data during the summer and winter periods in Thailand (October-March).
| Model | MAPE (2018) | MAPE (2019) |
|---|---|---|
| Seq2Seq-7 | 10.18±0.53 | 10.60±0.34 |
| Seq2Seq-14 | 10.41±0.43 | 10.51±0.41 |
| Seq2Seq-21 | 11.35±1.64 | 11.19±0.33 |
This project use Python 3.5.2 with the following dependencies:
A list of all required python packages can be found in requirement.txt
To install dependencies, run
pip3 install -r requirements.txt
For training your own model, we have provided a jupyter notebook, Training_example.ipynb, for handling the input and output data before feeding them to the model.
To launch a jupyter notebook server after installing all required python packages, run the command below in repo folder
jupyter notebook or jupyter notebook --port=7000 if you want to open jupyter server in port 7000
This will start your server at the URL localhost:7000. Copy and paste this URL into your web browser.
We provide the instruction for making UV forecast in Prediction_example.ipynb with example data.
Note: The model provided in this repository has an encoder input length of 1190 = 14 days * 85 time steps (10-minute intervals from 5am to 7pm) and a decoder input length of 85 = 1 day x 85 time steps.
If you want to make a forecast
- using input data from a different time period (e.g., other than 14 days prior to the forecast date), or
- for a different time period (e.g., other than the next day), or
- for a different weighted spectrum (e.g., other than anti-psoriasis),
then you have to train your own new model.