Time Series prediction of Solar Radiation using Transformer Neural Networks

Abstract

Global Solar Radiation (GSR) at any location is an essential requirement for the design of solar energy systems to be installed at that location. However, unlike conventional sources of energy the unpredictability of Solar Radiation is a major challenge in estimating and planning the usage of Solar energy systems. This paper presents the study, modelling and estimation of Global Solar Radiation (GSR) by considering solar radiation data as a time series function and using Transformer Neural Networks and Recurrent Neural Networks (RNNs). A Transformer Neural Network is the latest architecture in Machine Learning (ML) for handling sequenced data using attention mechanism. The results are also compared to other Machine learning algorithms handling time series data like Long Short Term Memory (LSTM) of RNNs. The results are compared using Root Mean Squared Error (RMSE) and Coefficient of Determinant (R <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> ) values. It was found that Transformer model provided comparable accuracy to LSTM model but was highly time efficient and stable while handling larger input and output data.