Abstract: A two-stage model for quantitative analysis of output uncertainty of photovoltaic power generation is proposed. In the first stage,the photovoltaic power for a period of time before the prediction day is used as the training sample,and the fuzzy entropy is introduced to quantify different weather types as the input. Then,based on the employing integrated empirical mode decomposition（EEMD）,the photovoltaic power time series is decomposed into multiple modal components,which are reconstructed into mesoscale and macro scale subsequences with Hurst index analysis. The BiLSTM based on attention mechanism（BiLSTM-Attention）is used to predict the two subsequences respectively. The error sequence corresponding to the mesoscale subsequence is corrected to obtain the point prediction results of photovoltaic power generation. In the second stage,the nuclear density estimation（KDE）method is used to predict the interval of photovoltaic power generation according to the error statistics obtained from the point prediction results in the first stage. The prediction interval coverage probability（PICP）at 95%,90%,85% and 80% confidence levels are obtained respectively.Taking the operation data of a photovoltaic power station in Northwest China as an example,the effectiveness of this method is verified.