To address the uncertainties in power flow states and the complexity of power optimization problems brought about by the high penetration of electric vehicles (EVs) and photovoltaics (PVs) in distribution networks

this paper proposes an active/reactive power optimization framework for high-penetration PV-storage-charging distribution networks based on a constraint-enhanced safe reinforcement learning method. Firstly

an EV charging station (EVCS) model based on transfer reinforcement learning is constructed to reasonably describe the charging and discharging process of multiple types of EVs considering demand response. Then

the active/reactive power optimization problem is formulated as a Markov decision process

and a mixed-integer programming-continuous double deep Q network (MIP-CDDQN) algorithm is proposed to solve it. This algorithm transforms the Max-Q problem for the continuous action value network into an MIP model while considering the real-time operational constraints of the distribution network

which ensures efficiency and security of the optimization strategy. Finally

simulation experiments on the IEEE 33-bus system demonstrate the effectiveness of the proposed MIP-CDDQN algorithm and its advantages in enforcing constraints and computational efficiency.