Abstract: Energy storage system connected with the end of distribution network， which is used for auxiliary services such as system voltage regulation， can effectively deal with the problem of voltage fluctuation caused by intermittent distributed renewable energies and the fluctuation of load demand. In this paper， the operation of energy-storage battery is modeled as a Markov Decision Making process. Considering its subsequent regulation ability， an intelligent control strategy based on deep reinforcement learning （DRL） is proposed. By embedding a Q deep neural network to approach the optimal action value， the problem of too large state space can be solved. The state vector composed of the state of charge （SOC）， the predicted output of renewable energy and the load level is used as the input of Q network， and the optimal discrete charge and discharge action is output， which is trained by replay strategy. Compared with the traditional method， the proposed method is based on learning without explicit uncertainty model， and the calculation efficiency is high. Finally， the IEEE 33-node distribution network system is analyzed by using MATPOWER in TtensorFlow， and the effectiveness of the proposed method is proved.

Key words: distribution network, battery energy storage, deep reinforcement learning（DRL）, voltage operation level