Posted on Leave a comment

Power Optimization using DRL for the Energy Internet


Xu, S., & Guo, S. (2024). Distributed Reactive Power Optimization for Energy Internet via Multiagent Deep Reinforcement Learning With Graph Attention Networks. IEEE Transactions on Industrial Informatics


Sagging voltages pose a danger to the stability of the Energy Internet (EI) with varying demands and dispersed power supply. Reactive power adjustment must be done quickly and precisely in these situations in order to preserve stability. To tackle this, the author proposes a reactive power optimization framework using multiagent deep reinforcement learning (DRL) for emotional intelligence (EI) during voltage sags. By employing real-time EI state awareness, this framework seeks to synchronize many reactive power compensation devices in order to guarantee voltage stability and introduces a Introduces a multi agent DRL to connect multiple reactive power compensation devices the proposed system validated through IEEE-9 bus system and an industrial zone.

Issues in existing method for managing voltage stability in EI system:

  • Reactive power methods lack swift response to fault circumstances
  • Traditional techniques may not efficiently handle abrupt voltage sags
  •  Limits exist due to focus on local bus system
  • Data abundance not fully utilized for optimal control
  •  Heuristic algorithms may not capture system dynamics accurately
  •  Simplified equations hinder optimal control utilization
  •  Methods too slow for online optimization support
  •  Challenges in simultaneous optimization of multiple devices
  •  Wide action space complicates arriving at improved solutions

For those above issues the author proposes a novel a novel framework for optimizing reactive power coordination based on graph attention networks (GAT) and multiagent deep reinforcement learning (DRL).

The expected reward for each action is calculated as:

 R_k(s) = Σ(i=k to n) γ^(i-k) * r_i

  • Where:
    `R_k(s)` is the expected reward of the kth action.
    `γ` is the reward attenuation coefficient.
    `r_i` is the reward of the ith reactive power compensation action.


Numerical simulations and field data from real systems construct the EI model. Voltage sags, disturbances, and faults simulate EI behavior. SVG controllers at nodes monitor conditions during sags. SCADA data from all nodes and PMU data from nearby nodes are utilized. High-frequency PMU data undergoes processing with GRU for sequential properties and GAT dynamically reveals EI topology. Combining GRU and GAT extracts EI features. A multiagent DRL model, based on A2C, is developed. Each SVG controller serves as an actor network agent. Cloud servers merge SCADA and PMU data to form a critic network. Agents utilize available state data for compensation decisions. EI simulation integrates agents’ compensation tactics. Reward systems evaluate compensation schemes. Reward values back-propagate networks during training. The trained multiagent DRL model adjusts reactive power online. It ensures voltage stability during sags.


Flowchart of proposed framework proposed from the study by  Xu, S(2024)


Results obtained during evaluation by IEEE-9 bus system from the study by Xu, S(2024)


  • The suggested framework efficiently improves voltage stability in the Energy Internet (EI) by combining multiagent DRL with Graph Attention Networks (GAT).
  • When compared to other techniques, it exhibits higher accuracy in reactive power adjustment
  • In order to optimize network setup for EI safety, future research will focus improving control mechanisms for catastrophic failure and natural catastrophes.




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 *