COMPARATIVE STUDY OF CONVOLUTIONAL NEURAL NETWORK AND LONG SHORT-TERM MEMORY NETWORK FOR SOLAR IRRADIANCE FORECASTING

Weather prediction is a problem that is one of the leading topics of research in Machine Learning. Since power generation from photovoltaic (PV) cells depends on the solar irradiance of each day, power generation prediction from PV cells can be related to the problem of weather prediction. Research related to Convolutional Neural Networks (CNN) applications is there, but the amount of work done on comparing the efficiency of it with various neural networks is close to none for forecasting. Long Short-Term Memory (LSTMs) networks is a kind of Recurrent Neural Networks (RNNs) that has predicted time series with exceptional accuracy. In this work, the efficiencies of CNNs and LSTM have been compared for global horizontal irradiance (GHI) forecasting in the short-term. LSTM networks are compared with CNNs in this application. The systematic evolution of the performance is done by adjustment of network parameters with different loss functions and training with satellite derived historical dataset. The proposed LSTM exhibits better results than CNN by training on data samples in 10 minutes interval. It also exhibits superiority to recent works for the hourly GHI forecast problem considered.