Abstract: In order to solve key problems like poor accuracy of current photovoltaic (PV) power forecasting models during extreme weather, lack of detailed analysis of weather conditions and absence of flexible adjustment methods, a photovoltaic power forecasting method that considering multiple scenarios and differentiated compensation strategy was proposed. In the method, a concept of mixed weather types was proposed, and adaptive classification of weather state scenarios was realized by constructing a clustering algorithm with an optimal evaluation function (COEF). Based on extreme learning machine, basic value prediction model was constructed, and compensation mechanism of multiple scenarios was clarified, and the multi-scale correction of the basic prediction value was realized by designing targeted error compensation models for different weather scenarios, so as to improve the prediction accuracy of the algorithm. Finally, multiple station real data from different regions with different climatic characteristics were selected for simulation testing. The simulation results show that compared with physical model and traditional machine learning algorithm, the proposed photovoltaic power forecasting method has better prediction effect under multiple time scales and multiple scenarios.