Solar Irradiation Forecasting by Long-Short Term Memory Using Different Training Algorithms

Deep learning nowadays is the hottest research topic with wide applications in various areas classifications, pattern recognition, solar forecasting, wind forecasting, etc. This paper presents the forecasting of one step ahead solar global horizontal irradiance (GHI) by long-short term memory (LSTM) deep learning network using three different training algorithms. The LSTM network is employed using three different training algorithms: adaptive moment estimation with momentum (ADAM), stochastic gradient descent with momentum (SGDM), and root mean square propagation (RMSprop). The 1 year of dataset is used for the training of the model, whereas monthly data is used to test the model. The two popular error metrics: root mean square error (RMSE) and mean absolute percentage error (MAPE) is considered to obtain the forecasting errors for each training algorithms-based LSTM model. The results prove that the LSTM with ADAM training algorithm outperforms the SGDM and RMSprop algorithm. The ADAM-based LSTM obtained the annual average RMSE of 68.62 (w/m2) and MAPE of 10.30%.