Abstract: Accurate day-ahead power prediction of photovoltaic generation is helpful to design future scheduling plan of power grid, reduce the impact of new energy generation on power grid and improve the accommodation rate. A day-ahead power forecasting method for photovoltaic generation is proposed based on the Boosting ensemble learning framework. Firstly, according to the characteristics of photovoltaic output which is mainly affected by weather, the meteorological factors with strong correlation are obtained through the Pearson coefficient, and the k-means ++ is used to cluster the total horizontal irradiance that is strongly correlated with photovoltaic generation power to obtain the similar daily datasets. And then, the extreme learning machine (ELM) is introduced into the Boosting framework to build the photovoltaic output day-ahead prediction model (B-ELMs). Finally, the validity of the model is verified using the operation data of real photovoltaic power stations. The proposed model shows good adaptability in the test process and has the highest decision coefficient (R²) of 0.9819. The experimental results show that due to the existence of ensemble learning framework, the B-ELMS can still provide accurate prediction results against the photovoltaic output curves with poor regularity and strong fluctuation under complex weather. At the same time, the B-ELMS has a faster convergence rate compared with deep-learning network, and can provide more accurate prediction results while maintaining faster training speed.

Key words: photovoltaic output forecasting, extreme learning machine, integrated learning, k-means, day-ahead forecasting