Photovoltaic Power Nowcasting Using Decision-Trees Based Algorithms and Neural Networks

Precise Photovoltaic (PV) power forecasting tools are needed to integrate PV into the new framework of the energy sector. Also, the related intermittent and random nature needs to be appropriately considered. In this context, this paper examines various machine learning algorithms used for one hour ahead forecasting of PV power production. Specifically, the performances of two Long Short-Term Memory (LSTM) recurrent neural networks, a Gradient Boost Machine (GBM) model and an Extreme Gradient Boosting (XGB) model, are compared. Six years of data are retrieved from the 81 kW PV power plant in the Savona campus of the University of Genoa and are used to train and test the algorithms. The performances of all the algorithms are compared over the original dataset, composed of meteorological variables linked with the PV production, a dataset using seasonal and trend decomposition (STL) of the meteorological variables, and some reduced datasets that mimic the situation in which some of the features are not available at least for some time steps. By comparing the results through different case studies, it can be justified that the XGB outperforms the other algorithms and the STL decomposition is helpful in increasing the performance of the model.