As the existing photovoltaic power prediction methods have low robustness under different weather conditions, we proposed a new method for the prediction of photovoltaic power system based on the unscented Kalman filtering (UKF) neural network. The method uses the unscented Kalman filter to update the weight of the neural network model in real time, and establishes two independent dual-input-single-output models with taking DC voltage and current as input and active power and reactive power as output. The experimental results indicate that the proposed UKF neural network model can accurately forecast the photovoltaic power, the best fit of active and reactive power are 97.3% and 94.2% respectively, and the method is robust to weather conditions.