Posted on Leave a comment

PV-Rich Networks: Voltage Management Boosting

Reference:

Chen, Y., Liu, Y., Zhao, J., Qiu, G., Yin, H., & Li, Z. (2023). Physical-assisted multi-agent graph reinforcement learning enabled fast voltage regulation for PV-rich active distribution network. Applied Energy351, 121743. https://doi.org/10.1016/j.apenergy.2023.121743

Overview:

Voltage violations stem from the intricate nature of active distribution networks, arising from the proliferation of distributed PV systems. Traditional systems encounter challenges in both efficiency and flexibility. To address this, a novel edge intelligence approach combines a graph attention network with multi-agent deep reinforcement learning. This innovative method effectively captures network dynamics and spatial correlations, optimizing voltage regulation. By incorporating an accurate physical model, it enhances learning speed. Demonstrated on IEEE 33-node and 136-node systems, this technique outperforms conventional approaches in both convergence and control effectiveness.

Limitations in other approaches for voltage regulation:

  • Traditional equipment like OLTCs and CBs effectively regulate voltage long-term but struggle in emergencies.
  • Electronic devices like STATCOM and SVCs respond quickly but often only handle reactive power.
  •  ESS offers flexible and rapid voltage regulation, but centralized control faces challenges in real-time adjustment.
  • Conventional MADRL algorithms struggle to efficiently train in dynamic distribution networks.

proposal and problem formulation:

To tackle this author proposes an edge intelligence method that uses a physical model that is correct for reference voltage regulation experiences, integrates graph attention network (GAT) into MADRL, and makes use of cloud-edge collaborative architecture and proposes a problem formulation for optimization formula as:

minPessj,t  ,Qsvcj,t   ∑tTjN,ij∈\Ꜫ lij,t rij

Here, Pessj,t   and Qsvcj,t   are the decision variables representing the active power supplied by ESS and the reactive power supplied by SVC, respectively. The summation is over the time slots T and network branches \Ꜫ.

Methodology:

For effective voltage control, the suggested system for voltage regulation integrates cloud and edge computing. The network is divided into sub-networks at the edge control level, each of which is represented as an agent and is responsible for controlling ESS and SVCs to provide dynamic voltage. By transmitting their judgments locally, these agents’ lower communication overhead by periodically transferring their experiences to the cloud for training. Centralized control takes place at the cloud learning level, where agents are taught with edge experiences that are saved in a replay buffer. Through the use of electrical distance and modularity, the network partitioning approach breaks down the concentrated difficulty into smaller, more manageable issues.

Markov game process:

A Markov game is used to model the multi-agent voltage regulation, in which agents interact with their surroundings and choose their course of action depending on observations. GAT-MASAC is a reinforcement learning technique that combines a MASAC for policy learning with a GAT to capture topological interdependence. Lastly, to improve the effectiveness of agent learning, a physical-assisted mechanism creates reference experiences based on an accurate physical model.

GAT-MASAC framework from the study by Chen, Y. (2023)

Comparision result of different approaches:

 

Case study result of different approaches compared with proposed GAT-MASAC from the study by chen,Y. (2023)

Conclusion:

  • In distribution networks with strong PV penetration, the suggested edge intelligence technology efficiently reduces power losses and mitigates voltage violations.
  • Graph Attention Networks (GAT) and MASAC together increase learning efficiency and agent adaptation to changes in network topology.
  • The method shows scalability for large-scale distribution grids, guaranteeing voltage stability and maximizing efficiency in operations.
  • Future studies should focus on improving ESS performance, investigating data-driven techniques for voltage control, planning the best possible resource allocation, and taking dynamic partitioning tactics into account.

Sakthivel R

I am a First-year M.Sc., AIML student enrolled at SASTRA University

I possess a high level of proficiency in a variety of programming languages and frameworks including Python. I have experience with cloud and database technologies, including SQL, Excel, Pandas, Scikit, TensorFlow, Git, and Power BI

Leave a Reply

Your email address will not be published. Required fields are marked *