【作者】 陈磊；

【导师】 王渝红； 魏巍；

【作者基本信息】 四川大学 ， 电气工程（专业学位）， 2022， 硕士

【摘要】 输电网在电力系统起着至关重要的作用,它实现了电能的大规模远距离传输,并且保证了电能传输的安全性与可靠性。而负载的不断增加和可再生能源的大量接入,给现有的输电网架带来了很多的问题。目前亟需一种既高效又能应对可再生能源不确定性的扩展规划方法。本文首先介绍了强化学习的基本理论,引出了深度强化学习深度Q网络(deep Q network,DQN)算法及其改进策略,然后基于DQN算法对传统输电网扩展规划问题进行求解,并验证了所提灵活扩展规划方法的有效性,最后提出了基于极限场景集与改进DQN算法提出了输电网鲁棒规划方法,用以求解风电接入下的规划问题。具体的研究内容如下:(1)对强化学习的基础马尔可夫决策过程进行介绍,并基于马尔可夫决策过程介绍了强化学习算法的基本原理,推导了最优动作值函数与最优策略的表达式;然后,介绍了Q-learning算法和结合了深度学习与Q-learning算法的DQN算法的基本原理,并提出了经验优先回放策略与竞争深度估值网络这两种DQN算法的改进策略。(2)基于DQN算法为多目标动态输电网扩展规划提供了一种新的解决思路。所提方法在规划时综合考虑了网架结构的经济性、可靠性和灵活性,同时在使用蒙特卡洛法时计及了可能发生的N-k故障以及设备故障的严重程度。利用DQN算法进行求解,基于DQN算法交互学习的特性来对线路的建设顺序进行了考虑,借助训练好的神经网络对方案进行灵活调整,通过仿真验证了所提方法的有效性。(3)针对风电接入下的不确定性问题,采用了极限场景集来应对其不确定性,提出了基于风电极限场景集的输电网鲁棒规划方法,并采用了经验优先回放策略与竞争深度估值网络两种策略来改进DQN算法,与传统规划方法相比更能适应风电出力的不确定性,通过仿真验证了其有效性。更多还原

【Abstract】 The transmission network plays a vital role in the power system.It realizes the large-scale long-distance transmission of electric energy,and ensures the safety and reliability of electric energy transmission.The continuous increase of load and the large-scale connection of renewable energy have brought many problems to the existing transmission network.There is an urgent need for an expansion planning method that is both efficient and can cope with the uncertainty of renewable energy.In this paper,the basic theory of reinforcement learning is introduced first,and the deep reinforcement learning deep Q network(DQN)algorithm and its improvement strategies are introduced.Second,the traditional transmission network expansion planning problem is solved based on the DQN algorithm,and the effectiveness of the proposed flexible expansion planning method is verified.Finally,based on the limit scenario set and the improved DQN algorithm,a robust planning method for the transmission network is proposed to solve the planning problem under the connection of wind power.The specific research contents are as follows:(1)The basic Markov decision process of reinforcement learning is introduced,and the basic principle of reinforcement learning algorithm is introduced based on the Markov decision process.The expressions of the optimal action value function and optimal strategy are deduced.Then,the Q-learning algorithm and the basic principles of the DQN algorithm combining deep learning and Q-learning algorithm are introduced.And two improved strategies of DQN algorithm,prioritized experience replay strategy and dueling deep Q network,are proposed.(2)Based on the DQN algorithm,a new solution is provided for multi-objective dynamic transmission network expansion planning.The proposed method comprehensively considers the economy,reliability and flexibility of the network structure when planning,and takes into account the possible N-k faults and the severity of equipment faults when using the Monte Carlo method.The DQN algorithm is used to solve this problem,and the construction sequence of the lines is considered by utilizing the interactive learning characteristics of the DQN algorithm.The scheme is flexibly adjusted with the help of the trained neural network,and the effectiveness of the proposed method is verified by simulation.(3)Aiming at the uncertainty of wind power coonection,a limit scenario set is used to deal with its uncertainty,and a robust planning method for transmission network based on the wind power extreme scenario set is proposed.The DQN algorithm is improved by adopting the experience-priority replay strategy and the competitive depth evaluation network strategy.Compared with the traditional planning method,it can better adapt to the uncertainty of wind power output,and its effectiveness is verified by simulation.更多还原