Vulnerability Analysis Of Optimal Power Flow Problem Under - PowerPoint PPT Presentation

Vulnerability Analysis Of Optimal Power Flow Problem Under Cyber-Physical Security Attacks Devendra Shelar and Saurabh Amin Massachusetts Institute of Technology INFORMS November 15 th , 2016

Vulnerable Electricity Networks: Key issues Two motivating attack models § In addition to reliability failures, power grids are increasingly vulnerable to cyber-physical security (CPS) attacks § Such CPS attacks can be modeled as bilevel optimization problems § We present two CPS attack scenarios California Sniper Attack § Dynamic Line Rating (DLR) Manipulation § Distributed Energy Resource (DER) disruption § Structural insights allow for greedy and efficient (approximate) algorithms Ukraine Cyber Attack

Related work § A. Verma, D. Bienstock: N-k vulnerability problem § Attacker disrupts generators or manipulates line impedances to maximinimize load shedding § DC Power Flow approximation § S.Wright et al.: Vulnerability Analysis of Power Systems § Attacker increases the line impedances to maximinimize § Loss of voltage regulation, OR Load shedding § Use both active and reactive power § R. Baldick, K. Wood, D. Bienstock: Network Interdiction, Cascades Use bilevel optimization models with outer problem as attack model, and the inner problem being optimal power flow (OPF) problem § Ensure demand is fully met while minimizing costs subject to generator, capacity, supply-demand balance, power flow constraints

DLR Manipulations in Transmission Networks Line capacity violations have cause cascading failures in the past, e.g., July 2012 blackout in India Bilevel problem (Stackelberg game) § Leader: Attacker compromises the DLRs using false data injection attack; § Follower: Defender’s economic dispatch solution is optimal for new manipulated system, but possibly infeasible for the old actual system. Problem statement: § Determine an optimal attack plan to maximinimize line rating violations

Solution Approaches Benders Decomposition (Kevin Wood et al.) § Alternately consider follower problem, with fixed attacker actions, and master problem with fixed defender actions § Sequentially generate Benders cut for the Master Problem until zero optimality gap § Results in systematic vertex enumeration of the inner problem Kuhn-Tucker Single-level reformulation (Bard et al.) § Apply KKT optimality conditions for the inner problem, and reformulate complementarity constraints § Use Branch-n-Bound techniques to solve the resulting Mixed-Integer Linear Program (MILP)

Insights 300 1000 u d 13 β 12 u 12 u d Actual DLR (in MW) 23 p 1 p 2 800 250 u s 13 , u s Demand (in MW) 23 f 12 Demand G 1 G 2 β 23 600 200 f 13 f 23 β 13 400 u 23 = u d u 13 150 23 200 u a 100 23 0 L 5 10 15 20 Time (in hours) § Attacker strategy, by and large, 300 exhibits a bang-bang policy u a 13 Manipulated DLR (in MW) u a 23 § Some DLRs are set to f 13 250 f 23 maximum 200 § Other DLRs are set to minimum (as long as feasible 150 operating point exists) 100 § Similar results hold for larger 5 10 15 20 Time (in hours) (118 node) testcase

Implementation of attack in Powerworld simulator 06410AE0 01 00 00 00 C0 65 49 09 00 00 00 00 00 00 00 00 06410840 01 00 00 00 A0 64 49 09 00 00 00 00 00 00 00 00 06410AE0 01 00 00 00 C0 65 49 09 00 00 00 00 00 00 00 00 06410840 01 00 00 00 60 65 49 09 00 00 00 00 00 00 00 00 06410AF0 00 00 00 00 00 00 00 00 00 00 00 00 00 FE 00 00 06410850 00 00 00 00 00 00 00 00 00 00 00 00 00 FE 00 00 06410AF0 00 00 00 00 00 00 00 00 00 00 00 00 00 FE 00 00 06410850 00 00 00 00 00 00 00 00 00 00 00 00 00 FE 00 00 06410B00 00 00 00 00 00 00 00 00 01 00 00 00 00 00 00 00 06410860 00 00 00 00 00 00 00 00 01 00 00 00 00 00 00 00 06410B00 00 00 00 00 00 00 00 00 01 00 00 00 00 00 00 00 06410860 00 00 00 00 00 00 00 00 01 00 00 00 00 00 00 00 06410B10 00 00 00 00 00 00 C0 3F E1 FA C7 42 E1 FA C7 42 06410870 00 00 00 00 00 00 C0 3F E1 FA C7 42 E1 FA C7 42 06410B10 00 00 00 00 9A 99 19 40 E1 FA C7 42 E1 FA C7 42 06410870 00 00 00 00 9A 99 99 3F E1 FA C7 42 E1 FA C7 42 (c) Pre-attack system state (safe). (d) Post-attack system state (unsafe).