Model predictive control

Quantum mechanics

Control (management)

Voltage

Process (computing)

Engineering

Control engineering

Operating system

Power (physics)

Scalability

Computer science

Distributed computing

Physics

AC power

Reinforcement learning

Database

Electric power system

Artificial intelligence

Decentralised system

Control theory (sociology)

Electrical engineering

IEEE Transactions on Smart Grid

infrastructure-enabled controls: centralized [1-3], local [4], distributed [5][6], and decentralized control [7][8][9].Centralized control strategy requires extensive communication links and can obtain global optimization results.However, it suffers from a computational bottleneck and the communications are not that reliable for today's distribution systems.Local control strategy only relies on local measurements and can react fast without high communication links.Due to a lack of coordination, only suboptimal solutions can be obtained and not all constraints are fully satisfied.The distributed control strategy achieves coordination between various control devices with limited communication links.The consensus-based methods are usually used, which are vulnerable to communication delays.Decentralized control combines the advantages of centralized and distributed methods by adopting zonal control and interzone coordination [10].To achieve decentralized or distributed control, network partition is a first step [11][12].In [7], a particle swarm optimization (PSO) algorithm is used for voltage regulation of each cluster.Although the PSO algorithm is easy to implement, it cannot guarantee a global optimum.To cope with the uncertainties of DERs and load demand, a robust optimization (RO) for distribution system optimization is proposed [9].Note that the solution of RO is achieved under the worst scenario, yielding conservative outcomes.Stochastic programming (SP) is also employed for uncertainty management [13].It depends on pre-sampling scenarios of uncertainty realizations, and therefore accurate knowledge of the distributions of random variables is needed.This is quite challenging to obtain.SP also suffers from heavy computational burdens.It is worth noting that accurate information on system topology and parameters is required for these methods.However, the model quality of the practical distribution systems is rather poor, especially under high penetration of the behind-the-meter PVs [14].The model assumption can be mitigated with the help of advanced machine learning-based control methods.They allow extracting knowledge from data to cope with uncertain patterns.The learned knowledge is scalable and thus can be exploited for optimization in new situations.Among these, the deep reinforcement learning (DRL) is widely used since it can learn

Data-Driven Multi-Agent Deep Reinforcement Learning for Distribution System Decentralized Voltage Control With High Penetration of PVs