Reactive Power Optimization Strategy Of Power System Based On Deep Reinforcement Learning

The application of artificial intelligence to reactive power optimization of power systems is the intersection of many hot research directions such as power system intelligence,practical application of artificial intelligence technology and power system optimization scheduling.It also highlights the contradiction of reactive power voltage regulation and voltage optimization in China It is difficult to control these problems that need to be solved urgently to provide new coping methods.Power system reactive power optimization can be achieved by adjusting reactive power control equipment,such as transformer tap setting,generator terminal voltage and reactive power compensation capacitor bank switching.Many traditional optimization control methods such as sensitivity ana lysis,quadratic programming,and linear programming require that the control variables are continuous and the objective function is differentiable,so these methods can easily lead to "dimensional disasters" and cannot be applied to large systems.Based o n the optimization principle of reinforcement learning,this paper studies a variety of reinforcement learning methods applied to the control strategy of power system reactive power optimization,aiming to achieve effective coordination of power system reactive power control equipment actions,reduce power system active loss,and realize the power grid Operate safely and efficiently.First,based on a brief description of the basic mathematical principles of reinforcement learning,the dynamic Markov process in reinforcement learning is applied to actual power system scenarios to achieve the establishment of mathematical models of reinforcement learning in power systems.The simulation results show that the reinforcement learning algorithm can solve the reac tive power optimization problem.Aiming at the traditional reinforcement learning,the value function iterative optimization method is not suitable for the problem of reactive power control of large systems,in this paper,deep neural networks are used to fit the value function,so that deep reinforcement learning is applied to reactive power optimization problems.Secondly,based on the deep deterministic strategy gradient algorithm(DDPG),a centralized control algorithm for solving reactive power optimiz ation problems is proposed,and the action information transmission mechanism of agents and power grid control equipment is designed.In order to accelerate the training speed of the neural network,a normalization layer is added between the fully connecte d layers to optimize the network structure of the algorithm.Analysis results of calculation examples show that the designed algorithm has higher solving efficiency than other intelligent algorithms;and it has better optimization effect.Then,based on the deep Q learning network(DQN),a distributed algorithm model for solving reactive power optimization problems is constructed.DQN is distributed in various reactive power control device agents to calculate the action value of control actions generated by the device.Then the agent uses the greedy algorithm to select the control action and execute it,so that the agent takes into account the exploration and development required in the reinforcement learning process.The discrete actions generated by the agent can be directly applied to the power system,which realizes the "end-to-end" control of the power system by the deep reinforcement learning agent.Finally,the active power transmission between the buses where the control devices are located is used to mine the reward data of each agent,the global reward data is obtained through the consistency theorem,and the multi-agent deep Q network(MADQN)model is established.The analysis results of the calculation examples show that the designed control method can effectively reduce the power system active power loss,and compared with the centralized algorithm,it reduces the time cost caused by the huge calculation amount.