The invention discloses an ultra-short-term photovoltaic generation power prediction method based on a self-learning composite data source. The ultra-short-term photovoltaic generation power prediction method based on the self-learning composite data source comprises the steps that data are input to enable parameters of an autoregression moving average model to be obtained; input data required by photovoltaic generation power prediction are input into the autoregression moving average model determined according to the parameters of the autoregression moving average model, so that a prediction result is obtained, post-evaluation is conducted on the prediction result, namely the error between a predicted value and a measured value is analyzed, and order determination of the model and estimation of the parameters of the model are conducted again if a predicted error is larger than an allowable maximum error. Key information is provided for new energy power generation real-time scheduling, a new energy power generation day-ahead plan, a new energy power generation monthly plan, new energy power generation capability evaluation and wind curtailment power estimation by predicting the photovoltaic generation power generated during photovoltaic generation. The ultra-short-term photovoltaic generation power prediction accuracy is effectively improved due to the fact the composite data source is introduced, and thus the on-grid energy of new energy resources is effectively increased on the premise that safe, stable and economical operation of a power grid is guaranteed.