A Blending Approach to Solar Energy Prediction

This work has the task to predict the solar energy production of 98 Oklahoma Mesonet sites based in weather forecasts provided by a 16x9 matrix of the Global Ensemble Forecast System (GEFS – Reforecast v2). Each GEFS system provides 15 weather variables and 11 forecasts splited in 5 daily intervals. We used raw data and dataset transformations to build 13 different datasets and trainned the models using gradient boosted regressor (GBR) with a 3 fold continuous crossvalidation technic to validate the results. Then the 13 models are ensembled using a Nelder and Mead mean absolute optimization algorithm to improved the performance. The method archieved the first place in the AMS 2013-2014 Solar Energy Prediction Contest and provides a very good prediction performance, since the mean absolute production error lies around 2.107MJ/m^2 daily, which represents a deviation of about 13% in the actual values.