In order to better realize the whole county photovoltaic roof intensive development, improve the economics of large-scale distributed photovoltaic grid roof ,analyze disadvantages of prediction method based on the meteorological data, explore the rule of the massive time-series photovoltaic power data , based on improved both short-term and long-term memory ,recursive neural network (LSTM) of the whole county comprehensive photovoltaic power prediction model is proposed. The model is used to predict a single photovoltaic power supply. The results show that the improved LSTM model is more accurate than other models, and it has higher overall prediction accuracy under the scenario of the whole county photovoltaic comprehensive power prediction. It can be used to predict the whole county photovoltaic power generation power, which is helpful to achieve a high proportion of distributed photovoltaic consumption.