Aiming at the problems that the selection of training samples is not reasonable and the input data of the model is not accurate enough, combined with practical engineering experience, a photovoltaic power prediction method based on sunny coefficient and multi-similarity matching is proposed. Firstly, the historical data are divided into sunny and non-sunny types, and two data queues are designed to store and update training samples. Then multi-similarity matching method is used to extract the measured power data of similar day by matching the predicted and historical data vertically and horizontally. Finally, an ARIMA time series prediction model based on the power of similar days is constructed. The analysis of practical application data show that the proposed method has achieved good prediction results in improving accuracy,and has practical significance.