Search

With photovoltaic (PV) penetration increasing, PV-output prediction has become a research hotspot. Due to the close correlation between PV-output fluctuation and weather conditions, PV-output prediction models often vary different weather types, while the historical/forecast weather types for modeling are mostly obtained from weather-service providers. However, weather-service providers generally have deficiencies in forecast accuracy, spatio-temporal resolution, and investment/operating costs. Based on the above, this paper changes the current acquisition way of the weather types, and proposes a framework of reversely determining weather types from historical PV-output data. First, the symbol-sequence histograms (SSH) are used to describe the PV-output volatility in a coarse-grained manner. Then, the SSHs are partitionally clustered and a classification rule for weather-types is proposed to label the historical weather types. Next, considering the chaotic characteristics of PV output, a prediction method combining phase-space reconstruction with an extremely learning machine based single-layer forward net is developed to predict the SSH. Finally, the day-ahead weather type is forecasted. Simulations were implemented on the weather-type classification and forecasting using a campus PV-system in East China. The PV-output prediction results show that, compared with weather information from a weather-service supplier, 75-day mean errors are significantly reduced by 15.55% (MAPE) and 12.69% (rRMSE), respectively. • A framework to reversely label weather type according to historical PV-output data. • The volatility of PV-output is quantified based on a coarse-grained manner. • A weather-type classification rule is proposed based on symbol histogram clustering. • Symbol-sequence histogram is forecasted using a phase-space reconstruction-ELM model. • The effect of the method is evaluated through the accuracy of PV-output prediction. [ABSTRACT FROM AUTHOR]