Research Article

Resiliency Assessment of Power Systems Using Deep Reinforcement Learning

Table 1

Recent studies on smart grid system security using RL and DRL.

ReferenceSystemMethodAttackRecovery actionAimLimitations

[16]Modified 9-bus systemDeep deterministic policy gradient (DDPG)Multiswitch attacks and false data injection (FDI) attacksReclose the transmission lines lost in the cyber-attack by optimizing the reclosing time.Reach the asynchrony in the power system by applying power blocking which will accelerate/decelerate the rotors of the generatorsOwing to its continuous action space, it will not be suitable for topological resilience studies
[17]IEEE 9, 14 and 30-bus systemsDeep Q-network (DQN)Data integrity attacksNo recovery actionEvaluate the delay-alarm error rates, false-alarm error rates, and detect-failure rates for the systemsDQN suffers from overestimation
[18]IEEE 30-bus systemDeep Q-network (DQN)Coordinated cyber physical topology (CCPT) attacksControl center can detect the line outage by using phasor measurement units (PMU) dataInvestigate the coordinated topology attacks in smart grid which combine a physical topology attack and a cyber-topology attack
[19]Wood & Wollenberg 6-bus system and IEEE 30-bus systemQ-learningSequential attacksAutomatic generation control (AGC)Identify the minimum number of attacks/actions to reach blackout thresholdQ-learning and SARSA techniques are limited to systems with small state-action space
[20]IEEE 14-bus systemSARSAFalse data injection (FDI), jamming, and denial of service (DoS) attacksNo recovery actionFormulation an online cyber-attack detection as a POMDP problem and propose a solution based on the model-free RL for POMDPs
Our workIEEE 6-bus systemDeep Q-network (DQN), double DQN, REINFORCE, and REINFORCE with baselineSequential attacksDisconnecting the faulted transmission linesEvaluating the resiliency of power systems against faults/attacks using DRLNeeds to investigate tabular methods such as Q-learning and SARSA to compare their performance with DRL methods