Day-ahead probabilistic photovoltaic power forecasting models based on quantile regression neural networks

This paper presents the results obtained in the development of probabilistic short-term forecasting models of the power production in a photovoltaic power plant for the day-ahead. The probabilistic models are based on quantile regression neural networks. The structure of such neural networks is optimized with a genetic algorithm which selects the values for the main parameters of the neural network and the variables used as inputs. These input variables are selected among a set of variables which includes chronological, astronomical and forecasted weather variables related to the location of the power plant. The forecasts correspond to quantiles of the hourly power generation in the photovoltaic power plant for the daytime hours of the day-ahead. The forecasts are obtained in the first hours of the day, allowing their use for preparing bid offers for the day-ahead in electricity markets