In order to reduce the prediction error of the photovoltaic system power and enhance the accuracy of the prediction results, this paper studies the short-term prediction method of the building photovoltaic system power, carries out data acquisition and correction, and establishes a normalized prediction matrix. Based on this, it constructs the BP neural network prediction model and realizes the short-term prediction of power by Mallat reconstruction processing. The final prediction results show that compared with the traditional multi-time scale forecasting group and the traditional distributed photovoltaic forecasting group, the final prediction errors obtained by the neural network photovoltaic prediction group are relatively small, and all of them are controlled below 0.15, it is shown that the prediction effect of the neural network photovoltaic prediction group to the system is better and it has practical application value.