Abstract: In a downlink multi-cell non-orthogonal multiple access system, power allocation is one of the key factors to determine system performance. Due to the non-convexity of the power optimization problem among multi-cell systems, it is very difficult to obtain the optimal power allocation. The power allocation algorithm based on deep reinforcement learning is proposed to maximize energy efficiency in this paper, which is simple and efficient. The algorithm takes the deep Q network as the action-state value function, system energy efficiency is directly set as a reward function, which optimizes channel power allocation and maximizes system energy efficiency. The simulation results show that the algorithm of proposed scheme is more effective than the weighted minimum mean square error, fractional programming, maximum power and random power algorithms in achieving higher system energy efficiency. The scheme also has better performances in algorithm calculation complexity, convergence speed and stability.