Research On Voltage Control Strategy Of Active Distribution Network

Building a new power system with a large amount of renewable power generation is an important way to achieve "carbon peak" and "carbon neutralization" goals.Distribution network is an important way for renewable energy to connect to power grid.However,owing to the uncertainty and intermittent of renewable energy,their integrations bring challenges for the safe and economic operation of distribution network.The voltage rise issue caused by two-way power flow is the most prominent problem,which is also the main factor restricting the wide access of renewable energy.Active distribution network can deal with the voltage problems caused by high proportion of renewable penetration by actively scheduling various controllable devices to guarantee the safe operation of distribution network.This dissertation focuses on the voltage regulation of active distribution network with the support of National Key Research and Development Program of China “Research on Key Technologies of planning and stability control of weakly interconnected hybrid renewable energy system”(2018YFE0127600),the main contents are as follows:1 This dissertation proposes a local control strategy of PV inverter for the voltage regulation of distribution network based on multi-agent deep reinforcement learning algorithm features centralized training and decentralized execution.To deal with the fast voltage fluctuations caused by the rapid variation of PV generation,a coordinated local control strategy is proposed by casting the coordination control of PV inverters into Markov games,which model each PV inverter as an actor-critic-based reinforcement learning agent.The centralized and decentralized execution mechanism is introduced to formulate a coordinated control strategy utilizing historical/prediction data.The critic augmented with information of other agents contributes to the formulation of a cooperative strategy.The attention model utilized in the proposed method further help each agent attend to the information that is most related to its reward duting the training process.When the training is completed,only the actor network of each agent is kept.The actor networks can inform real-time decisions based on local information.2 This dissertation proposes a distributed voltage regulation strategy of distribution network based on network partition and multi-agent deep reinforcement learning.To deal with the voltage fluctuations caused by the intermittent of renewable energy generation and achieve real-time optimization of distribution network,this dissertation proposes a distributed control framework based on multi-agent deep reinforcement learning.The electrical distance index is first defined based on the voltage-reactive power sensitivity.Then the spectral clustering method is applied to partition the whole network to several sub-networks to decouple the correlations between the nodes and identify the control areas of the distribution network.Then each sub-network is modeled as an intelligent agent,all of which are trained in a centralized manner.Owing to the structural difference between action function and evaluation function,the agent can learn a cooridnated control strategy during the centralized training process.When finish the training process,the proposed method can achieve coordinated scheduling of control devices based on regional information and inform fast decisions to deal with voltage fluctuations caused by rapid variation of PV generation.3 This dissertation proposes a data-driven distributed voltage control strategy based on surrogate model and multi-agent deep reinforcement learning algorithm.To deal with the uncertainty of physical model of distribution network,a physical-model-free distributed voltage control framework based on surrogate-model-enabled multi-agent deep reinforcement learning algorithm is proposed.Sparse Gaussian process regression algorithm is applied to learn the complex mapping relationship between the active and reaction power injection and voltage magnitude of each node utilizing recorded historical data.The posterior model is treated as a surrogate of the distribution network.The electrical distance is defined according to the voltage-active power and voltage-reactive power sensitivity,based on which the network partition is achieved while considering the regional voltage regulation capability.The coordinated scheduling of multiple sub-networks is cast to Markov games and solved by a multi-agent deep reinforcement learning algorithm features centralized training and decentralized execution.During the training process,the surrogate model is embedded in the environment of the Markov games.It can provide reward signal for agents,which learns by continuous interaction with the surrogate model.The proposed method can achieve coordinated scheduling of multiple sub-networks and provide real-time decisions without the requirement of accurate physical model of distribution network.4 This dissertation proposes a data-driven two-timescale control framework for the voltage regulation of distribution network.Since the characteristics of mechanical devices and power electronic devices are different,it is difficult to achieve the coordinated control of these devices.To this end,a two-timescale control strategy is proposed for the coordinated control of two kinds of assets in this dissertation.The scheduling of mechanical devices such as on-load tap changers and switched capacitors is modeled as a Markov decision process,which is solved by a deep reinforcement learning algorithm considering the future uncertainty and long-term impact of the actions.To fully utilize the fast respond capability of power electronic devices,a distributed control method is used for the scheduling of PV inverters by dividing the whole network to several sub-networks and applying multi-agent deep reinforcement leraning algorithm.Sparse Gaussian process regression method is utilized to learn the complex nonlinear mapping relationship between power injection of nodes and voltage magnitude from historical data.The upper-level agent and lower-level agetns are trained in a centralized manner according to the reward signal calculated by the surrogate model to develop a coordinated control strategy.The two level agents are trained concurrently with information exchange to achieve systematical coordination of the two-timescale devices.The devised training mechanism and the utilization of surrogate model enable the proposed method to achieve coordinated control of multiple kinds of devices and the real-time scheduling of power electronic devices without the requirement of physical model.