Abstract: Aiming at the problem lacking of surface solar radiation(SSR） data when forecasting the photovoltaic(PV） power generation, a PV power generation forecast method without SSR information is proposed. First, the astronomical radiation feature is added on the original data. Considering that this feature may still not meet the prediction accuracy requirements, output data under similar meteorological data is introduced as another augmented feature. Second, the method of feature weighting is proposed for the similarity algorithm, and the weight coefficients of each feature are obtained by the synthetic scoring of three scoring methods of Spearman, maximum information coefficient, and the importance of random forest features through the D-S evidence theory, so as to extract the meteorological data and further improve the ac‍curacy of the similarity algorithm. Finally, the similar day search results and similar hour search results under Euclidean distance and cosine similarity methods are compared. The study of a PV power plant in Ningxia shows that the weighted Euclidean distance similar hour search results obtain the optimal forecast performance, that the average accuracy rate and qualification rate of the four seasons reaches 0.889 and 0.944 respectively, promising a relatively good forecast result of PV output without the SSR information.

Key words: surface solar radiation, photovoltaic power generation, forecast, D-S evidence theory, similarity algorithm, weighted similarity algorithm