Abstract: Aiming at the problem that traditional PV hosting capacity (HC) assessment methods relied too much on detailed physical models and led to the limitation of computational complexity, a PV hosting capacity assessment method combining the attention mechanism and model-free deep reinforcement learning was proposed. A deep neural network (DNN) was constructed to predict the node voltages and was combined with the Soft Actor Critic (SAC) algorithm for HC assessment, and the interactive attention mechanism adaptively focused on the intrinsic correlation between state-action pairs to improve the estimation accuracy of the Q-value and the training stability. The case study was unfolded in a real power distribution network and was compared with the physical modeling method based on DIgSILENT and the model-free method without attention mechanism, and the experimental results show that its total average absolute error is 0.0452, total root mean square error is 0.0618, average absolute percentage error is 5.93%, the coefficient of determination is 0.9305, and the maximum voltage deviation is around ±3V, which verifies the validity of the method.