Abstract: The electricity generation of photovoltaic power plant is affected by various factors such as external weather condition and equipment performance loss, showing the strong volatility and stochasticity. Therefore, the accurate prediction of PV output is crucial to the operation and management of PV power plant and the scheduling operation of power grid. To solve the shortcomings of traditional fuzzy C-means clustering（FCM） that cannot determine the number of clusters autonomously and the Euclidean distance in classifying high-dimensional data, in this study, the adaptive factor and cosine distance as the sample classification index are integrated with traditional FCM, leading to the set of historical sample clusters with the highest similarity to the data to be predicted. Finally, a shortterm PV output portfolio prediction model based on improved FCM and long-short term memory（LSTM） neural network is innovatively established. The applied results in a PV plant in Yunnan show that the proposed method has the better classification of historical samples and the higher prediction accuracy of power generation than other prediction models, which verifies the effectiveness and superiority of the method.