An active distribution network voltage control using artificial intelligence

In this chapter, two network partition-based decentralized voltage control methods are proposed. The conclusions are summarized as follows:A distributed coordination control method is first proposed for distribution system Volt-VAR control, considering both PV inverters and SVCs. The spectral clustering algorithm allows for a partition of the large distribution system into several sub-networks from a voltage control perspective. Then, the control of each sub-network is formulated as the MGs and solved by an attention-based MATD3 algorithm. The proposed method is centralized-trained and distributed-implemented and can easily be used for real-time voltage regulation. Compared with centralized control, the proposed approach mitigates issues, such as communication bottlenecks and privacy concerns. Compared with other distributed control methods, only local information is needed without communications between agents. The proposed method can be adapted to the flexible network partition requirements of the operator rather than the typical MADRL algorithm. Comparative results with several other existing model-based and data-driven methods demonstrate that the proposed method can achieve 96.4% optimality based on local information while also considering uncertainties. However, the model-based approach could not achieve satisfactory outcomes in the case of rapid variations in PV outputs. A novel MADRL framework is introduced for voltage regulation in distribution systems with a high PV penetration, featuring centralized training and decentralized execution without physical models. The SPGP is first leveraged to build a surrogate model that learns the mapping relationship between the active and reactive power injections and the voltage magnitude of each node using few-shot recorded data. This surrogate model is further integrated with the MADRL to assist in the formulation of a coordinated control strategy. In particular, the voltage regulation problem is cast into the MADRL framework by partitioning the whole network into several sub-regions, considering both the regional voltage regulation ability and the electrical distance. Each sub-region is treated as an agent and solved by the DRL algorithm. All agents are trained in a centralized framework to learn the coordinated control strategy guided by the reward given by the surrogate model. The proposed method can achieve real-time scheduling using only local information. Comparative results demonstrate that: (1) the proposed decentralized control strategy can achieve similar performance as the centralized one; (2) the performance of the suggested model-free approach closely matches that of one relying on a flawless physical model; (3) the control strategy can be taken in real-time with the usage of the battery storage system to mitigate the influence of violent PV fluctuations.The future works include the development of a new control method that can coordinate the inverters and utility-owned equipment, which is a two-timescale control problem. We will also propose a meta-learning-based MADRL algorithm to deal with topology changes in the distribution networks.