High-precision photovoltaic power prediction is one of the key technologies for the efficient use of solar power generation. To satisfy the requirements of high-precision photovoltaic power prediction, an ultra-short-term photovoltaic power generation power prediction method based on multi-feature extraction is proposed. Firstly, the relative position relationship between the sun and photovoltaic array is analyzed, and the irradiance mechanism model of the inclined surface of the photovoltaic (PV) module is established. Secondly, considering the influence of the temperature change of the PV module backsheet and attenuation of module efficiency on the photovoltaic power generation, the deep neural network and convolutional neural network are used to extract the temperature characteristics of the module backsheet at each time and slow time-varying characteristics of the photovoltaic array, respectively. Finally, considering the timing correlation of photovoltaic power generation, the long short-term memory neural network is used to extract the dynamic time series features of the data. Based on the collected historical power generation power and historical meteorological data, the input of the prediction model is constructed via the established multiple feature extraction modules, and the prediction value of photovoltaic power generation power is obtained in the next 15 min to 4 h via the multi-layer neural network. Considering a photovoltaic power station in Zhejiang province as an example for ultra-short-term photovoltaic power generation power prediction, among the prediction results of each model, the prediction accuracy of the photovoltaic power generation power prediction model based on multi-feature extraction is higher than that of other comparison models. Furthermore, the monthly average accuracy can reach 97.13%, which verifies the effectiveness of the proposed prediction method.