Solar Irradiance forecasting using Recurrent Neural Networks

Abstract

Solar irradiance being the chief constituent of the solar power extraction is dominated by the atmospheric conditions. Prediction of irradiance data is highly sought after in the field of forecasting and predictive maintenance. For this purpose various machine learning methods are being used to improve the accuracy of the forecasted value. This paper aims at prediction of solar irradiance using Recurrent Neural Networks (RNN) using Long Short Term Memory (LSTM) architecture. Using different combinations of input in the supervised learning method the accuracy for single as well as multiple time steps are determined. The results are shown in the form of evaluation metric as well as the forecasted values and actual value comparison. It is seen that for single time step prediction the LSTM RNN puts out highly accurate values but error for higher time steps prediction accumulates in a compounded manner. It is also observed that using time based models along with the inputs increases the accuracy of the forecasted values. © 2022 IEEE.