Using the historical generation and weather data (hourly) between the dates 01.01.2019 – 30.11.2021, forecast the generation (hourly) for the period 01.12.2021 – 31.12.2021.
I used CatBoost for final submission but realized that I should have spent time on other algorithms like tabnet or NN. I will also add other algorithm solutions to the repo.
Here is my solution:
- I have read a lot of papers (articles, thesis etc.) because the most importance features (from my perspective) come from domain-specific informations.
- I mostly focus on external data like sun position, irradiance etc. You can use this library: https://pvlib-python.readthedocs.io/en/stable/index.html
- Another extraordinary thing was Campbell-Norman (1988) method. I calculated irradiance using Effective Cloud Cover and GHI-DHI-DNI, you can see the method here: https://pvlib-python.readthedocs.io/en/stable/forecasts.html?highlight=campbell#cloud-cover-and-radiation
- My feature selection method (LOFO) is coming from: https://github.com/aerdem4/lofo-importance
Some of my references (the tip of the iceberg):
William F. Holmgren, Clifford W. Hansen, and Mark A. Mikofski. “pvlib python: a python package for modeling solar energy systems.” Journal of Open Source Software, 3(29), 884, (2018). https://doi.org/10.21105/joss.00884
Campbell, G. S., J. M. Norman (1998) An Introduction to Environmental Biophysics. 2nd Ed. New York: Springer.
Wang Y, Feng B, Hua Q-S, Sun L. Short-Term Solar Power Forecasting: A Combined Long Short-Term Memory and Gaussian Process Regression Method. Sustainability. 2021; 13(7):3665. https://doi.org/10.3390/su13073665
Ertekin, S. (2020). Solar Power Prediction with an Hour-based Ensemble Machine Learning Method . Hittite Journal of Science and Engineering , 7 (1) , 35-40 . DOI: 10.17350/HJSE19030000169
Daniel O, Kubby J. Feature Selection and ANN Solar Power Prediction. Research Article. Journal of Renewable Energy, 2017. https://doi.org/10.1155/2017/2437387