Power grid transmission line defense method based on multi-agent deep reinforcement learning

Abstract

The invention discloses a power grid transmission line defense method based on multi-agent deep reinforcement learning, and the method comprises the steps: training a deep neural network of each agentthrough employing a preset training method based on the environment state, the attack behaviors of the agents and the rewards of the agents; according to the deep neural network and the environment state obtained through training, enabling each agent to autonomously determine an attack behavior, and obtaining an optimal attack line set of each agent after attack resources are used up; repeating the above steps for many times, and obtaining an optimal defense line set according to the optimal attack line set of each agent to perform key defense. According to the invention, the performance lossof the power grid caused by multi-stage cooperative transmission line attack can be effectively reduced.

Description

technical field [0001] The invention relates to a power grid transmission line defense method based on multi-agent deep reinforcement learning, and belongs to the cross technical field of smart grid security and artificial intelligence. Background technique [0002] The large-scale interconnection of power grids realizes the optimization of large-scale resource sharing. While achieving significant economic benefits, the highly coupled characteristics of information and physical systems in the power grid have brought new challenges to the safe and stable operation of the power grid. When a cyber or physical attack is launched, it is possible that the transmission line will be disconnected, which will trigger a series of chain reactions and cause large-scale power interruptions in the power grid, such as the 2012 blackout in India. Therefore, it is of great significance to identify and defend these transmission lines that may trigger a chain reaction and lead to grid disassem...

AI Technical Summary

Problems solved by technology

In fact, research results based on reinforcement learning methods show that given the same attack resources, multi-stage attacks can cause greater losses than single-stage attacks

However, this method only considers small-scale grid

When the scale of the power grid gradually increases, the state space will increase, and the action space will also increase sharply, which will make the existing methods no longer applicable (such as unable to converge)

Although the challenge of large state spaces can be overcome with deep reinforcement learning methods, the exponential growth of action space with the number of attack stages makes single-agent deep reinforcement learning methods not scalable

Images