Research On Reactive Power Optimization Method Of Distribution Network Based On Deep Reinforcement Learning

Reactive power optimization of distribution network refers to the rational distribution of reactive power by adjusting the reactive power compensation device of the system to realize the local balance of reactive power,which can reduce the network loss and improve the power quality.However,with a large number of electric vehicles and distributed generation connected to the system,modeling and power flow calculation become more and more difficult.Moreover,the previous reactive power optimization algorithm applied to large distribution network has the disadvantages of long operation time,insufficient memory and easy to fall into local optimization.At the same time,there is also a large amount of operation data in the distribution network,which can be applied to the distribution network.Therefore,based on big data technology and deep reinforcement learning technology,aiming at the current optimization problems,a new reactive power optimization method of distribution network is proposed.The main research contents are as follows:(1)Firstly,the classical IEEE-37 node distribution network model is transformed based on OpenDSS platform,and the OpenDSS-MATLAB joint simulation platform is constructed,and the optimization method based on the platform is proposed;Then the improved model is optimized by using particle swarm optimization algorithm and joint simulation platform,and the optimization results are analyzed,which establishes the foundation for the next research.(2)A two state agent based on deep confidence network is constructed.Firstly,in order to extract the operation characteristics of massive data in distribution network,this paper proposes a new method to extract data characteristics: fusion feature method.Then,in order to make full use of the historical data and provide rewards for the next deep reinforcement learning algorithm,a two-state agent is constructed,which takes the operation state and control strategy of the distribution network as the input,and the network loss and voltage deviation as the output respectively.From the function mapping relationship,the relationship between input and output is learned by using the deep confidence network,and the network loss agent and voltage deviation agent are trained.Finally,the improved IEEE-37 node model is used to train the network loss agent and voltage deviation agent model,and it is proved that the two agent models are effective.(3)A reactive power optimization method based on deep reinforcement learning algorithm is proposed.In this paper,the deep reinforcement learning algorithm is used to optimize the reactive power of distribution network.Using the above-mentioned two-state agent,the optimization problem is transformed into a deep reinforcement learning agent,and a multi-agent system is constructed.Then,using the Double Deep Q-learning Network(DDQN)algorithm in deep reinforcement learning,a reactive power optimization method based on DDQN is proposed.The dual state agent provides rewards to obtain the optimal strategy.Finally,the DDQN model is trained through an example,and the optimization effect of the model is analyzed to prove that the method proposed in this paper is effective.(4)The adaptability of this method is analyzed.Firstly,this method is compared with Deep Q-learning Network(DQN)algorithm and Particle Swarm Optimization algorithm.The results show that this method does not depend on distribution network model and parameters,can realize online reactive power optimization,and the optimization time is greatly reduced.Then,the reactive power optimization of this method is carried out under different photovoltaic permeability and different communication conditions,and its optimization effect is discussed respectively.It is verified that this method also has strong robustness and generalization ability under the distribution network with low perception and high permeability.