AC Power Short-term Forecasting of a Thin-film Photovoltaic Plant Based on Artificial Neural Network Models

Solar and PV forecasts for time horizons ranging from a few minutes ahead to several days ahead are generally significant for planning the operations of power plants which convert renewable energies into electricity. With the increasing penetration of photovoltaic (PV) power systems into the grid, the problems caused by the fluctuation and intermittence of PV power output are gaining interest. The power output fluctuations impact the power system’s stability. For this reason, an accurate forecast of PV production is necessary to consider PVs a reliable energy source.Different techniques can be used to generate solar and PV forecasts, depending on the forecast horizon—very short‐term forecasts (0–6 h ahead) perform best when they make use of measured data, while numerical weather prediction models are essential for forecast horizons beyond approximately 6 h.The aim of this study is to evaluate models for PV AC power short-term forecast using different artificial intelligence-based techniques and using ahead values of solar radiation and ambient temperature as data sources for forecasting. The data refer to a 1-kWp experimental micromorph silicon modules plant located at ENEA Portici Research Centre in southern Italy. A large dataset consisting of data measured every 5 min and acquired from 2006 to 2012, is used for the training/testing of the artificial neural networks (ANNs) proposed here. The AC power production evaluation is based upon data measured by a commercial inverter used for plant connection to the grid. This kind of inverter allows acquisition of operative data during their functioning hours, which are usually the central hours of the day. Therefore, when commercial inverters are used to acquire data, the use of ANNs is the best method for forecasting.In order to verify the effectiveness of the forecast data, measured and predicted data have been compared and the errors have been calculated by means of the relative mean bias error, the relative root mean square error and the correlation coefficient. Experimental data are reported to demonstrate the potentiality of the adopted solutions and to compare the different techniques proposed here. Furthermore, an algorithm that allows classification of a day as variable, cloudy, slightly cloudy or clear has been used for verification as forecast uncertainty depends on the meteorological conditions.