Photovoltaic Power Forecasting with Ensemble of Learners: Large Test Case from PV Plants in Italy, Zambia and Australia

Photovoltaic power forecasting is nowadays a very active research topic both for industries and academia [1- 10]. In this paper a hourly day-ahead forecasting approach for photovoltaic (PV) power is described, which is based on a number of machine learning techniques independently trained and then combined in a cooperative ensemble fashion. The results of the single techniques have been in depth analysed and compared, resulting in a superior performance of the cooperative ensemble. A strong contribution of this work is the application to a very large number of PV plants (42 at the date of this paper), for an overall installed nominal power over 110 MW in Italy and over 140 MW in Zambia and Australia. Another contribution is the large historic dataset dimension comprehending years of hourly data from 2014 to 2019, allowing in most cases to test the algorithms in large periods, taking into account seasonality and time varying factors. Moreover the large test case leads to a number of different PV technologies (monocrystalline silicon, polycrystalline silicon, thin-film amorphous silicon and flexible amorphous thin-film silicon) and installations (ground and roof mounting, axis tracking). Examples of studies in the literature that include so large test cases are scarce. This work extends the results of a previous work of the authors [11], where a smaller test case was used and tracking installations were not present.