Abstract: Aiming at the difficult problems of reactive power and voltage control in power grids with high proportion of distributed generation access, an improved reinforcement learning method is proposed, constructs a multi-scale feature fusion and adaptive exploration mechanism, establishes a three-level cooperative control architecture, transforms the optimization problem into a Markov decision process, designs a strategy-value dual network and integrates multi-constraint cooperative verification. Simulation tests are carried out on the IEEE 33-bus system. The results shows that the voltage deviation is reduced to 2.1%, the network loss rate is 2.13%, the convergence time is shortened to 186 s, the voltage over-limit rate under high permeability conditions is less than 1.18%, and the safe operation duration under full constraints reaches 23.7 h, which significantly improves the control accuracy, real-time performance and robustness.