Abstract: Under the background of large-scale integration of wind power,the voltage overlimit problem of wind farms is becoming more and more serious.To alleviate such problems and improve the safety and economy of wind farms,volt/var control methods have attracted extensive attention.However,traditional volt/var control methods rely on accurate wind farm models. With the fast construction of wind farms,limited operators,and variable environments, many wind farms lack well maintained models, making it difficult to implement effective volt/var control.In this paper, a model-free volt/var control method based on improved deep reinforcement learning is proposed. The optimal control strategy is obtained by intelligently mining the online control samples.To improve the efficiency and stability of existing deep reinforcement learning based methods,stochastic strategy and flexible constraint technology are introduced to avoid early local optimality.Numerical experiments have verified the effectiveness and superiority of the proposed method, which can adapt to the unmaintained wind farm models and provide efficient voltage control service.