PV power prediction in a peak zone using recurrent neural networks in the absence of future meteorological information

Abstract

As the majority of daily PV power outputs is mostly obtained in a peak zone around noon, hourly PV power output prediction in a peak zone is considered as an essential function for more sophisticated operations of PV facilities. However, the prediction of PV power output in a peak zone is a challenging problem since meteorological information is continuously changing and difficult to obtain for a particular area. In addition, due to only using the meteorological information observed in the morning to estimate PV power outputs around noon, the input features which are utilized as a shorter horizon from the horizon of the prediction are making the problem even more complex. Therefore, this study proposes two PV power output prediction model by using long short-term memory (LSTM) and gate recurrent network (GRU). In particular, unlike the previous methods, the proposed models attempt to understand the hidden sequential patterns of PV power outputs based only on the information captured in the morning without utilizing future meteorological information observed around noon during training. The experiment results using a real-world dataset indicate that the proposed models perform better PV power prediction in the peak zone than conventional models.