Under special conditions such as rapid changes in solar irradiance, local shading or poor solar irradiance, the traditional method will cause a photovoltaic (PV) power generation system to oscillate near the maximum power point (MPP), and the tracking speed will be slow. In serious cases, the PV power generation system will exhibit severe power fluctuations, which will lead to divergence. To solve this problem, a maximum power point tracking (MPPT) control method for PV power generation systems based on reinforcement learning (RL) is proposed in this paper. The RBF neural network weight coefficient in RL is adjusted online and in real time to adapt to the output control signal, so that the active power output from PV power generation can be quickly tracked to the MPP. A PV power generation system was constructed in MATLAB/Simulink, and a comparison with the conventional RBF neural network control method was performed. Results show that the RL control can make the PV power generation system converge to the MPP stably and accurately in a shorter time. At the same time, the PV power generation system can output the maximum active power.