This work is a effort for develop a machine learnig model for predict the solar radiation.

In all solar industries the correct prediction of solar radiation is a key on design and operation proccess. Due the stochastic nature of solar radiation the existing empirical and mechanistic models are good aproximations but with considerable errors.

The model developed in this work was trained with a dataset with meteorological and geographical data of all country (Mexico). The data was obtained from PV-GIS trhough a scraper method and proccesed with the library PV-LIB developed by Sandia National Laboratories

• mining_solar_info

  • This folder contains all logic for scraping the data and save the resulted datasets.

• model

  • model/feature-engineering: Is a collection of notebooks with the feature engineering proccess.
  • model/model-selection: Is a a csollection of notebooks with the model selection procsess.
  • model/ghi: This folder contains all logic for tunning all best models, compare the result and generate plots and metadata of each result.

Link to Kaggle

The resultant model with the best performance was a:

XGBR (Xtreme Gradient Boosting Regretion)

• SantaAna (Sonora Mexico)

• R2 = 0.9467821940386125
• RMSE = 76.9138282703692
• Training time: 7.2580 seconds

• R2 = 0.9354960908553472
• RMSE = 83.21137350309252

• Real vs Predicted

• Scatter Plot