Aiming at the congenital shortcomings of traditional multi-resistance intelligent algorithms to deal with the problems of accurate modeling of heterogeneous photovoltaic power forecasting, such as the lines multi-impedance parameter constraints, lower fluctuations, and line loss analysis easily falling into local extremes, a prediction model for short-to-medium-term photovoltaic power generation is proposed based on improved depth deterministic policy gradient (DDPG). Firstly, by introducing multi-agent mechanism and considering the parameters involved in the power generation system as independent active agents, constructed is a global optimal collaborative control system oriented to power generation process parameter information sharing with social attributes. Then, the battery energy storage power can be adjusted independently and accurately and the power grid output power can be automatically and optimally predicted by using the improved DDPG algorithm. Finally, based on the Tensorflow open source framework, the model efficiency was simulated under the Gym torcs environment, and a model heterogeneous photovoltaic power generation network was used as the performance evaluation carrier to verify the rationality the model.