Univariate short term forecasting of solar irradiance using modified online backpropagation through time

In this paper an attempt has been made to predict the solar irradiance values for multiple look ahead time predictions with time intervals as small as fifteen minutes. The recurrent neural networks in the past have been implemented on datasets with an interval of at least 30 minutes. The recurrent neural network was trained using backpropagation through time and the prediction was done using only the past solar irradiance values. A sliding window implementation of the network was achieved and the training time for over 20000 data points was less than 5 minutes which is an improvement over the execution time of other deep learning architectures. The online form of back propagation through time was implemented with the modification that the network took into account both the past mistakes and the current direction to which it is moving. The performance of the proposed network was tested comprehensively using two years of data and it was compared with the performance of persistence model and the normal recurrent network. The modified backpropagation network outperformed the baseline models for different time intervals. The proposed network also showed improvement over results computed for dataset with 15 minutes interval which was not achieved by the earlier state of the art architectures.