Machine learning system for solar energy prediction from meteorological data

Helios is a machine learning system to predict solar energy production from meteorological data. Solar energy provides several environmental benefits over traditional fossil fuels, but the amount of energy it yields considerably varies, depending on outside weather conditions. Energy companies therefore need to forecast and adapt to energy production to ensure that the optimum ratio of renewable and fossil fuels are available at all times.

The model employs a Tikhonov Regularization model, trained on a dataset from the American Meteorological Society Committees on Artificial Intelligence Applications to Environmental Science, Probability and Statistics, and Earth and Energy. Using a grid search over log space, the hyperparameter alpha was tuned to yield optimum performance.

The model predicts the total daily incoming solar energy at 98 different Mesonet(https://www.mesonet.org/) stations in Oklahoma. Weather data was harvested from the NOAA/ESRL Global Ensemble Forecast System (GEFS). The model was trained on historic data from 1994 to 2007. The system's accuracy is tested on data from 2008 to 2009.

Using the mean absolute error metric to quantify accuracy, the algorithm performed at MAE 2560862.16. This outformed classical algorithms often used as benchmarks, such as spatially weighted Gaussian mixture models (4019469.94) and Catmull-Rom spline interpolation (MAE 2611293.30).

In conclusion, Tikhonov Regularization is a computationally efficient and effective method to predict future solar energy yields from meteorological data. Future work that could increase further the performance of this model could include simplifying feature selection methods (including PCA and Boruta's adaptation of the Random Forest classification algorithm), or training the model over a larger hyperparameter space.

Built with

Try it out