Accurate photovoltaic power generation prediction plays an important role in the safe operation of photovoltaic power generation system. However, due to the instability, intermittent and randomness of solar energy, the existing short-term prediction models of photovoltaic power generation have problems of large prediction error and low generalization ability. Therefore, a Hybrid Neural Network (A-HNN) and attention mechanism for short-term forecasting of distributed photovoltaic power station is proposed. The temporal and spatial characteristics of data are extracted by Residual LSTM and dilated causal convolution, and an improved hybrid neural network model is obtained by adding attention mechanism to enhance feature selection. According to the characteristics of the time series of power generation data, the time series data with daily cycle are selected. Finally, compared with other recent models, the results show that the hybrid model can greatly improve the accuracy of photovoltaic power generation prediction under the same conditions. This work is supported by Ningxia Natural Science Foundation (No. 2020AAC03034) and Scientific Research Innovation Project of China Western First-class Universities (No. ZKZD2017005).