Visual-Based Reinforcement Learning for Voltage Regulation and Attack Mitigation in Distribution Networks

doi:10.1109/TSG.2025.3612285

Visual-Based Reinforcement Learning for Voltage Regulation and Attack Mitigation in Distribution Networks

2025-09-19

0

PRE

AI

DOI：10.1109/TSG.2025.3612285

Original

Original request for help

Save

Abstract

En 中文

The growing integration of distributed generators introduces increased vulnerability to cyber threats in modern power distribution systems, especially false data injection attacks (FDIAs) targeting communication network. This paper presents a novel real-time voltage regulation framework that enhances both system security and operational efficiency under FDIA scenarios. A time-frequency visual (TFV) model is first developed to identify FDIA types and accurately detect their duration in real-time. Based on the detection outcomes, a redirected control mechanism is employed to correct compromised control signals. These components are embedded into a Markov decision process (MDP) that guides voltage regulation strategies. To solve the MDP, an attention-based multi-agent soft actor-critic (AMS) algorithm is introduced, leveraging attention mechanisms to enhance decision-making under complex and uncertain environments. Integrating the TFV and AMS modules, a visual-based reinforcement learning (VRL) approach is formulated for robust and adaptive voltage control. Extensive simulations conducted on a modified IEEE 33-bus system using real-world synchrophasor data demonstrate that the proposed method effectively contains voltage deviations within 0.02085 p.u., significantly improving the resilience of the distribution network against cyberattacks.

Keywords:

False data attack detection

time-frequency visual

voltage regulation

multi-agent reinforcement learning

distribution systems

Visual-Based Reinforcement Learning for Voltage Regulation and Attack Mitigation in Distribution Networks

Abstract

Journal

Organization