Deep learrning based PV power forecasters in python for different time horizons

Abstract

The major points worked on throughout this report are: achieving accurate fore- caster with less complexity and computational cost, using the minimum available data set for training and reaching the farthest possible span in the future. For the aim of developing forecasters in this work, then RNNs and DL were employed with the use of the python programming language for their modelling. A data set of GHI recordings collected during January 21, 2011, through March 4, 2012 and from December 20, 2012, through January 20, 2014 is used to compare the above DNN based models for three different time spans. Moreover, various evaluation metrics such as MAPE, RMSE, r and R2 have been used for the assessment of the models to explore their performance when spanning different time horizons such that each one has a specific training samples. The obtained results have showed that the AE LSTM is the most efficient and less sensitive to the number of training samples.