Abstract: The massive grid connection of distributed photovoltaic (PV) systems has exacerbated the fluctuation of nodal voltage magnitude and the degree of three-phase unbalance in distribution networks. Traditional algorithms fall short in solving the three-phase unbalanced distribution network voltage control problem in terms of solution speed and optimality. Against this backdrop, this paper introduces a physical-model-assisted deep reinforcement learning method tailored for three-phase unbalanced distribution network voltage control. Initially, the method primarily considers the reactive power regulation capability of PV inverters and energy storage's active power regulation capability, constructing a distribution network voltage control model that integrates nodal voltage magnitude deviation and voltage three-phase unbalance. Subsequently, to address this three-phase nonlinear coupled control issue, the proposed physics-model-assisted deep reinforcement learning algorithm incorporates the physical characteristics of energy storage into the reinforcement learning algorithm, ensuring that the state of charge of the energy storage does not deviate from the feasible domain during offline training and online decision-making processes. Ultimately, the effectiveness of the proposed method is validated using the IEEE three-phase 123-node test case.