The invention discloses a BP neural network photovoltaic power generation system power prediction method based on a similar day. After quantization coding of weather influence factors, a historical day, which has higher similarity with a prediction day, is selected according to a similar day selection principle; through power generation data and weather data of the similar historical day, and with meteorological data of the prediction day being combined, a training sample set is formed to carry out training on a BP neural network; and during training, a weight adjustment algorithm adopts a gradient correction method with additional momentum and variable learning rate being combined, thereby improving model convergence speed, reducing the probability that a model falls into locally optimal solution and ensuring prediction model precision and stability.