Adversarial Attack for Deep Reinforcement Learning Based Demand Response

Document Type

Conference Proceeding

Date of Original Version

1-1-2021

Abstract

Demand response (DR) has been considerably used to improve the stability and efficiency of power system. Among the existing optimization and control models for DR programs, deep reinforcement learning (DRL) can directly learn optimal management policy from the input data and provide dramatic improvement in capability. However, DRL is vulnerable to adversarial attack where small perturbation is added into the input data. To investigate the vulnerability of DRL-based DR under adversarial attack, this paper formulates this problem as a nonlinear integer programming problem. A Particle Swarm Optimization (PSO)-based method is proposed to solve this problem and generate the attack vector. The proposed approach can identify the most vulnerable data points. To the best of our knowledge, this is the first time that adversarial attack for DRL-based DR is investigated. The performance of the proposed method is evaluated under extensive simulation analysis.

Publication Title, e.g., Journal

IEEE Power and Energy Society General Meeting

Volume

2021-July

Share

COinS