Research On Dispatching Method Of Highly Volatile Power System Based On Deep Reinforcement Learning

With the construction of New Power System dominated by new energy,the proportion of variable renewable energy generation such as wind and light continues to increase,and the high volatility attribute of the power system becomes increasingly prominent.This places higher demands on the flexibility and adjustment ability of the system.System scheduling is an important means of effectively regulating resource allocation,and it plays a crucial role in achieving the safe and stable operation of New Power System.Therefore,based on the deep reinforcement learning method,the article focuses on the randomness and large-scale characteristics of the power system,and conducts research on the scheduling problem of high volatility power system.Based on this,the article proposes a new intelligent scheduling method suitable for high volatility power system.The method provides theoretical support and method reference for the scheduling research of high volatility system.The specific research content of this article is as follows:Firstly,in order to study the macro impact of volatility on the power system,the article defines a high volatility power system as one system with variable renewable energy,random loads,and various disturbance factors(such as random attacks),where the system has a huge line scale,node scale,power and load access volume.By analyzing the impact of the randomness and large-scale characteristics of New Power System,the challenges faced by the current high volatility power system scheduling problems are summarized.By comparing the advantages and disadvantages of various modeling methods in the application process of power system scheduling,the modeling advantages of reinforcement learning in high volatility power systems are clarified,and the article proposes improvement directions for the reinforcement learning method.Secondly,in order to cope with the impact of bilateral randomness and randomness attacks of the system,the paper proposes a high-dimensional feature statistical encoding strategy based on deep neural networks.This effectively characterizes the high-dimensional features of high volatility sources in the power system.Once again,the article establishes a model for scheduling problems in power systems from the perspective of deep reinforcement learning.The article quantifies and defines the problems of substation switches,line switches,and generator output in the power system,and the article designs the state spacespace,action spacespace and benefit function of the agent.Then,the article improves the accuracy of agent in the power system scheduling process by embedding high-dimensional vector features of high volatility sources.The article uses the multi-layer strategy distillation method,which effectively reduces the state space of the agent,and provides a hot start for the actual training of the agent.On the basis of the Soft Actor-Critic algorithm,the article adds the action of rescheduling.Based on this,an agent training strategy suitable for high volatility power system is proposed.Finally,the article comprehensively considers factors such as generator output and load changes at various time scales,and adds the cost of rescheduling.The average cost during the power system scheduling process is used as an evaluation indicator to evaluate the effectiveness of agent in different volatility scenarios.To solve the scheduling problem of the high volatility power system,the article uses the Grid2Op power scheduling simulator to conduct simulation experiments on the proposed system’s intelligent scheduling method.The result shows that,compared with the approximation algorithm,the method in this paper can not only ensure the long-term effectiveness of the system,but also make the system more economical in the scheduling process when dealing with real-time scheduling problems of high volatility power systems.