【摘要】 分布式能源系统凭借其高效、环保、经济、可靠、和灵活等特点成为中国能源未来发展的重要方向。目前中国的很多分布式能源系统经济效益较差，主要原因是能源系统没有良好的运行策略。提出了一种基于深度强化学习的分布式能源系统运行优化方法。首先，对分布式能源系统的各个设备进行数学建模；其次，深入阐述了强化学习的基本原理、深度学习对强化学习的结合原理及一种基于演员评论家算法的分布式近端策略优化(distributed proximal policy optimization, DPPO)算法流程，将分布式能源系统运行优化问题转化为马尔可夫决策过程(Markov decision process, MDP);最后，采用历史的数据对智能体进行训练，训练完成的模型可以实现对分布式能源系统的实时优化，并对比了深度Q网络(deep Q network, DQN)算法和LINGO获得的调度策略。结果表明，基于DPPO算法的能源系统调度优化方法较DQN算法和LINGO得到的结果运行费用分别降低了7.12%和2.27%,可以实现能源系统的经济性调度。更多还原

【Abstract】 Distributed energy system has become an important direction of China’s energy development in the future by virtue of its high efficiency, environmental protection, economy, reliability and flexibility. At present, many distributed energy systems in China have difficulties in making ends meet, the main reason is that the energy system has no good operation strategy. A method of distributed energy system operation optimization based on deep reinforcement learning was proposed. Firstly, mathematical model was established for each device of distributed energy system. Secondly, the basic principle of reinforcement learning, the combination principle of deep learning and reinforcement learning and a distributed proximal policy optimization(DPPO) algorithm based on actor critic algorithm were discussed in detail, and the operation optimization problem of distributed energy system was transformed into Markov decision process(MDP). Finally, the historical data were used to train the agent, and the model completed by training can realize the online optimization of the distributed energy system, and the scheduling strategy obtained by deep Q network(DQN) algorithm and LINGO was compared. The results show that compared with the results of DQN algorithm and LINGO, the operation cost of the proposed energy system scheduling optimization method based on DPPO algorithm is reduced by 7.12% and 2.27% respectively, which can realize the economic scheduling of the energy system.更多还原