In terms of model-free voltage control methods

when the device or topology of the system changes

the model's accuracy often decreases

so an adaptive model is needed to coordinate the changes of input. To overcome the defects of a model-free control method

this paper proposes an automatic voltage control (AVC) method for differential power grids based on transfer learning and deep reinforcement learning. First

when constructing the Markov game of AVC

both the magnitude and number of voltage deviations are taken into account in the reward. Then

an AVC method based on constrained multi-agent deep reinforcement learning (DRL) is developed. To further improve learning efficiency

domain knowledge is used to reduce action space. Next

distribution adaptation transfer learning is introduced for the AVC transfer circumstance of systems with the same structure but distinct topological relations/parameters

which can perform well without any further training even if the structure changes. Moreover

for the AVC transfer circumstance of various power grids

parameter-based transfer learning is created

which enhances the target system's training speed and effect. Finally

the method's efficacy is tested using two IEEE systems and two real-world power grids.