Power Grid Contingency Analysis with Machine Learning: A Brief Survey and Prospects Sam Yang *Center for Advanced Power Systems, Florida State Universit Bjorn Veagensmith', Deepika Patra! Tallahassee, Florida 32310 ‘College of Computing, Georgia Institute of Technology, Atlanta, GA 30332 Hdaho National Laboratory, Idaho Falls, Idaho 83415 Email: syang@caps.tsu.edu, Bjorn. Vaagensmith @inl gov Abstract—We briefly review previous applicat learning (ML) in power grid analyses and introduce our ongoing effort toward developing a generative-adversarial (GA) model for fast and reliable grid contingency analyses. According to our review, the persisting limitation of traditional ML techniques in {rid analyses i the need for an exhaustive amount of training data for model generalization and accurate predictions. G els overcome this limitation by frst learning true data dist from a small waning set, from which new samples ass true data are generated with some variations. Subsequently, GA models ean transfer learn or supersgeneratize with increased ‘that is, accurately predict 1 — (Je + 2) contingencies uall n —& training set and generated » — (41) data. ‘fort between Mdaho National Lab and Florida State strives to develop a zero-shot and deep learning-based ency analysis tol, named ‘Contingency’ Analysis Neural. Network (SCANN), by leveraging. the aforementioned advantages of GA models. The basic architecture of SCANN stems from the Latent Encoding of Atypical Perturbations network combined with an adversarial network, and itis designed to generate imbalanced power flow data from learned true data distributions for prediction purposes. Here we also introduce the abstract concept of resllence-chaos plots, a new resilience characterization tool proposed to complement SCAN by aiding in the assesment of large amounts of high-order contingency predictions. Index: Terms—comtingency analysis, machine learning, power rid, resilience, SCANN. of machi 1. INtkonucTION |HE massive 2003 northeastern blackout, which resulted in economic losses estimated between 7-10 billion dol- Tars, was a formidable event prodding at resilience deficiens within the North American power grid [1]. In response to the disaster, the North American Electric Reliability Corporation produced the TPL-001-0.1 standard outlining conditions for When systems are required to withstand an n— 1 contingency without any service interruption, North American utilities 10- day are required (o comply with this standard; hence, present- day power outages are mostly due to n — 2 or greater conti gency events caused by extreme weather conditions, operator error, or malicious actions. Cyberattacks targeting n — 2 oF greater contingencies are also emerging as severe threats to ‘modern power grids. In March 2019, for instance, the Western “This work was support though he INL Laboratory Directed Research & Development (LDRD) Prigran ander DOE Idaho Operations Ofce Const DE-ACIT-OSIDI4SI7, 978-1-7281-8693-1/20/831.00 ©2020 IEEE transmission grid experienced a denial-of-service attack which caused a grid cyber disruption [2], [3]. The capability to quickly identity and understand higher order contingencies has become of interest to the power engineering community as a reasonable means to improve system reliability and resilience. ‘Apart from investigating very specific scenarios of inter est, utilities do not evaluate system-wide n — 2 or greater contingencies due to the high computational cost required to simulate all possible grid configurations—an exhaustive n—k analysis requires nl/(n — 6)! iterations assuming that sequence matters. The Texas 2000 bus test case, for exam- ple, contains 537: different components (i., power lines, busses, and transformers), requiring 5373 different power flow simulations to identify every vulnerability within the system given ann — 1 contingency case. n — 2, —3, and n—4 contingencies, however, will require approximately 2.6 X 10%, 5 x 10°, and 6.5 % 10° times the number of x — 1 power flow simulations, respectively. As a result, utilities have no decisive way to effectively preempt all possible n 2 oF greater ‘vulnerabilities owing to the exponentially inereasing computa- tional demand—they are rather forced to adopt an emergency response-type fault management for fixing problems without prior knowledge. The DC power flow approximation is widely used to quickly ‘compute power flow across a grid while neglecting all non- linear characteristics [4]. The method is typically employed during early grid design or assessment stages as well as to provide reference power flow data for evaluating newly proposed methods. Other numerically efficient algorithms and methodologies exploiting graph theory, matrix properties, or Stochastic processes have been proposed t0 address the hur= dies associated with the combinatorial nature of contingency analysis [5]-[10], Machine learning (ML) techniques have also been adopted to leverage available real power flow data and to rely on statistical methods without the need to solve intricate nonlinear equations [11]-{28]. In fact, ML-based power flow predictions are now becoming more popular, mostly due to their practicality and flexibility, which are essential to cope with the growing demand for data-driven real-time power man agement framework such as in digital twin-hased applications 129) ‘This paper focuses on ML techniques applied in contin gency analyses; in particular, we discuss the previous appli 119 Auorzed ices use ive io Lula University of Technolgy. Downloaded on Apl 90 202 a 1708-17 UTC fam IEEE Xplore Resbichons acycations and limitations of such techniques and introduce @ joint effort between Idaho National Lab (INL) and Florida State University (FSU) to develop Smart Contingency Anal- ysis Neural Network (SCANN), a novel zero-shot and deep Teaining-based contingency analysis (ool. SCANN is currently under development based on the Latent Encoding of Alyp- ical Perturbations (LEAP) network [28] complemented by an adversarial network, and it is designed (© provide new imbalanced power flow data from latent space For both (raining and testing purposes as in cransfer learning. To aid in data visualization, resifience-chaos plots are also introduced as means (o quickly characterize how tolerant a system is to n—k contingencies. ‘The rest of the paper is organized as follows: In Section I we review representative works corroborating the promising potential of ML-based contingency analyses in a chronological order and discuss the areas deserving further scrutiny. In Section III, we introduce SCANN and provide an overview of its technical novelty and envisioned outcomes. The resilience- chaos plots complementing SCANN are then introduced in Section IV followed by our concluding remarks in Section V. IL, PREVIOUS STUDIES ‘We down-selected previous works by the number of eta tions forthe brevity of our discussion, Table I summarizes the representatives works reviewed herein, from which we pick a few and discuss in greater detail below. ‘Semitekos and Avouris [18] combined ML. techniques with statistical modeling to produce experimental data for eval- uating the “nature” of contingencies. In particular, the au- thors implemented decision trees, generalized nearest neigh: bor, BayesNer [30], and multilayer perceptron (MLP) and determined the threshold of features for every contingency. Semitekos and Avouris observed thatthe best feature selection algorithms tend to point out to the most significant power transmission indices and/or voltage profile indices in a given power network, and the nearest neighbor outperformed other ML algorithms considered in their study Rudin et al. (21) employed support-veetor machines. to predict abnormal grid-related events and to produce failure ‘and vulnerability rankings based on historical New York City ‘power grid data. In addition, feeder failure rankings, cable, joint, terminator, and transformer rankings, feeder Mean Time Between Failure estimates, and manhole events vulnerability rankings were obtained with the model to support decision ‘making. Inthe paper, the authors validated the model, demon- strated its practicality, and discussed the challenges associated with the use of historical data for predictive modeling. ‘To be specific, the authors suggest utility data to possess the following properties for predictive modeling: First, the data should be as clean as possible, eg., unique identifiers should be used for each grid component. Second, the properties of the old component (and its surrounding context if it is used to derive features) must be recorded before replacement to retain the common properties. Verma and Niazi [22] proposed a supervised learning ap proach for fast and accurate power system security assess ment and contingency analysis. The model consisted of two independent deep feed-forward neural networks (DFFNNS) which predicted vollage-reactive power performance index (PIVQ) and tine MYA performance index (PIMVA). The ‘model adapted a resilient backpropagation scheme for updat- ing its weights [31] and tested the effectiveness ofthe proposed methodology on the IBEE 39 bus New England system at different loading conditions corresponding to @ single tine outage [22]. The classification accuracy for both PIVQ and PIMYA was around 99%: Donnot et al [241 proposed the guided dropout method and ‘combined it with a DFFNN for fast power system security analysis. Unlike the conventional dropout, the guided dropout relied on a deterministic approach wherein cach neuron in the dropout layer was activated and deactivated based on the state of the comesponding grid clement, c.g, power line disconnection, offline generator, etc. As a result, the network become more robust and correlated to clementary rid topology variants. The model was first trained on n — 1 ‘contingencies and was then generalized to x —2 cases without retraining, which is also referred to as super-generatization In Ret. [25 the same authors also employed the trained mode! to rank both n— 1 and » — 2 contingencies in a decreasing order of presumed severity based only on power line thermal capacities calculated in nm — 1 offline simulation data. ‘The ‘proposed model outperformed the DC approximation method and was also evaluated for scales up to 1000 power lines, ie, French High Voltage power grid. Kim et al. [26] applied a supervised graph convolutional neutral network (CNN) for predicting an optimal load-shedding ratio that prevents transmission lines from being overloaded under line contingency. The grid topology information such as connectivity was convoluted over the neural network (NN) as an adjacency matrix, The authors validated their model ‘against standard TEEE cases and compared against a classical NN and a Tinear regression mest, according to which GCNN outperformed the others by an order of magnitude, Similarly, Donon et al. [27] adapted GCNN and proposed a. graph NN solver which generalized the grid topology information to achieve zero-shot learning, Not only the proposed model ‘converged faster than a fully-connected NN. but it was also ‘capable of predicting power low in grids it was never trained fon. more accurately than the DC approximation, ‘More recently, Donon et al. [28] proposed Latent Encoding of Atypical Perturbations (LEAP) networks for system iden tification. Although the underlying concept of LEAP nets is similar to that of graph NN and ResNer [32], they do not require an explicit description of the grid topology under analysis, ic, adjacency matrix is not required. This is a remarkable advantage over existing ML-based contingency analysis models as lack of such knowledge is a practical problem grid operators encounter often. ‘The authors showed the superior performance of LEAP nets over ResNer, fly connected NN, and the DC approximation in predicting power 120 Auorzed cans use veo Lula Univers of Technolgy. Downloaded on Apl 90 202 a 1708-17 UTC from IEE Xplore Resbichons seyTABLET ‘SUMMARY OF REDRESENEATIVE PAPERS ON MAC AUTHOR, YEAR Refac t a, 1999 (13) ML TECHNIQUES Radia Hass Function NN Srivasav ot a, 2000 [18] Hybrig NN Nisa ota 2008 161 MUP Semitetos and Avouris, 2006 [18] Decision ess, generalized ros nein, BayesNr, tnd MLP ‘Singh and Seivastava, 2007 [19] sale NN ‘Swarup, 2008 [20] INN with patornreogsition Rudin ota, 2012 211 sv DPENN with cosiiot ack: propagation Verma and Nia, 2012 (22) Donnot ot a, 2018 (28), (28) DPENN with guided dropout kim ta. 19 25) Graph CNN Donon ca, 2019 (27) Graph NN Donon eal, 2020 1281 LEAP nats HIGHLIGHTS “Proposed a RBENN model o explit is nonlinear mapping eapsilitics for Cstimating line low and bus volage Following 2 conungency. » Architecture Tested tv RBFNNS, one for extimating the line flow and theater for bis sollge magni.» Teste om CIGRE.Tl-bassyom + Proposed a hybrid NN fr voltage srening un ranking based onthe votiage ‘perfomance index. + Adopted a iter module (DFFNN) to distinguish erica ontingcncies fom ponerical ones and wanking mess hsm af HENNY for furor classiction of ential contingencies.» Tested on IF 30-us and 75s Ian stems ‘+ Prosehad an MLP model wih diverging-backward sequential fete Section algorithm,» Tesined the NN with relent hackprpasation, «Tested fon THEE S7-hus sytem. “Combined sattical modeling and MIL techniques oem the “nau” of Cantngoncy. = The bet fet selection algorithms pint otto the me ‘Spniicant power tansmission indies andor voltage prof indies.» The cues neighbor outperformed othe ML algorithms considered in heir ty. “Proposed a Cascade NN for line flow seeing and ranking. + Adopted the hyo NN propose in Ro. [JAD with angular distance hase elustring or Feature selection.» Tesed a [tus ri + Enployed a. NN with patiem recognition for contingency analyses of rower system + The NN inca} sages whore th firs lage clsiod fn opening point as scody-aate scout or smsccue, he sont ape lassiiod the Sealy state soeure a tansienly secure or insecure, and Ube last sage claiie!the steady state secure and trie slate secure sales into dynamically cure or isceure. + Tesed on Dns ye, + Assessed proactive msntenance programs for NYC elctcal grt eit ‘used on historical grid data and ML.» Vr the accuracy and Mex of the medctand discussed the challenges associated with the use of isoicl Gat or predictive redoing + Proposed a supervised DFENN model for dstrmining power system soca Satu along ih contingency screcing and tanking» The achtetue Featured two DFFNNs and resent Rackprogration to prdict PIVO and PIMA. + Achieved 99% classification aos.» Tete on TEE, 39-05 Now England system, ‘+ Proposed a DEFNN with guided dropout to capture the elementary grid topology vatanis for super generalization. + Validated against tho Trench ‘Etia High Volage power gid dita The model oupecfonnod the DC spproxinaion method» Tested for upto 1000 oes. “Adapted CNN and graph propeics to exploit power grid topology infor ration white pricing an epimal lo shaking ratio,» Valted agains Standard IEEE cases,» The model outperformed wadtonal NN and linear repression mel Tested for ap to TT nes, + Proposed a graph NN solver based on GCN and gencalized grid topology information to achieve zer-shot learning.» Capable of prodicting powes ow in grid wax never tained on moe acuraely than the DC apprnimain, Tested foe up 0 110 noes + Proposed LEAP nets whose underying arcitetre is simi to that of {raph NN and ResNer witout the need Toran explicit exciton of the gs topology. « Veried superior performance of LEAP nets oer ResNer and or ropresetative modes in predicting power ow. * LEAP nets outperformed ‘ther models in swperenerlzine the tained model wo new. imbalanced test data. + Tesed for up to 192 substations (Pench us igh vl grid Towlome). ow in grids of different scales as well asin super-generalizing the trained model 1o new, imbalanced test data as in transfer learning, ‘The growing interest toward ML-based grid modernization and resilience enhancement is further reflected by industrial ‘and government initiatives. The North American Energy Re- siliency Model promoted by the U.S. Department of Energy Ofice of Electricity (OF), for example, plans to adopt ML to optimize the utilization and security of the energy sector [33] Similarly, the Grid Resilience & Intelligence Platform (GRIP) project administered by Stanford Linear Accelerator Center and Berkeley Lab aims to develop and deploy a suite of novel ML-based power flow and grid analysis tools to anticipate, absorb, and recover from severe contingencies [34], GRIP 121 Autborzedlicensed us lmted to: Lulea University of Technology. Downloaded on Apd 30-2024 at 1708-17 UTC from IEEE Xplre. Resbcions aly.Nearest Neighbor svm. Decision Trees FENN Fig. 1 Implomenution porcanage of each ML techigue for power rid oningency analysis where RNN-rcustent neutal netoek, FENN-feed- forward neral network, and SVM-suppot-vecior machines. Aces eieved from Gongle Sehr via woh sping exploits ML techniques for distribution grid resilience with the goal of promoting 10% reduction in the economic costs of power outages by 2025 [34]. The Blectrie Power Research Institute (EPRI) has also demonstrated tremendous potential of ML in accurately idenlifying critical events on the power arid by analyzing massive amounts of data. In particular, EPRI investigated the applicability of supervised and unsupervised ML techniques and proposed novel hybrid frameworks that combined them (see Refs. [35], [36). Fig. I clasifes ML techniques discussed earlier by their im- plementation percentage based on the scholarly data retrieved via web scraping. We used the following common keywords to serape these articles from Google Scholar without date re- striction: contingency analysis, machine learning, power flow, and power system. The retrieved articles were then filtered to eliminate review papers, textbooks, and duplicates from our analysis. According to Fig. 1, around 60% of previous studies \wo scraped employed decision trees and SVM, most probably due to their popularity before NN gained traction. Readers may ao refer to Ref. [37] for a comprehensive review of ML-based power system analysis. The quest for a fist and reliable grid contingency analysis and selection algorithm, requiring less training data is evident scconding to our literature review. Furthermore, the immense ‘potential and robusiness of ML over traditional deterministic tnd stochastie methods have catalyzed the scientifle commu- nity to quickly adapt and develop novel MIL-based conting analyses tools. Our literature review, however, alludes (0 the Jimitation persisting in conventional ML-based tools includ ing the super-generattzarion approach—these methous require extensive amounts of precomputed power flow data and are thereby limited to 2 problems in practice. A more practical ML-based contingency analysis algorithm, capable of solving 2 or greater contingency problems and down selecting * Actual numbers in Fig. | may vary due fostons. seca and filtering imper- critical events, is still needed to dismiss the need for large amounts of precomputed training data. In an effort to address this challenge, INL and FSU have partnered to develop Smart Contingency Analysis Neural Network (SCANN). IIL, SMART CONTINGENCY ANALYSIS NEURAL NETWORK Fig. 2 depicts a notional SCANN architecture devised based. (on a conditional generative adversarial network (CGAN) [381 and the LEAP net [28]. In SCANN, the generator is repre- sented by a LEAP net pre-trained on 1 and » — 1 contingency data, Subsequently, the generator predicts 1 — (k + 1) power flow for k > 1 and given grid topology (supplied to the Tatent space as in Ref. [28]), which are then “discriminated” against precomputed 1 — (+1) data, The power flow data consist Of real power, reactive power, vollage, and current (For power lines only) predicted for each component 1 in the grid. The discriminator receives the grid topology as @ conditional input, and it feeds back the generated samples classified as real to the generator for subsequent predictions, e.g...» — (k +2), n= (E-+3), and so on, ‘The major difference between the LEAP net (28) and SCANN is the exploitation of grid topology embedded into the latent space for the generation of new power flow data sam- ples. By leveraging LEAP net's super-generalization, we are designing SCANN to self-train with higher order contingency data, SCANN is considered successtul if it implicitly learns the tue date distribution; hence, we evaluate the samples it generates based on maximum mean discrepancy (MMD) [39], MMD verifies if two sets of samples—one from the generator and one from the true data distribution—come from the same distribution, and it has been observed to be more informative than either generator of discriminator loss [40]. In SCANN, we compute the squared difference of the statistics between the two sample sets elated bya kemel Given a kemel Ks X x YB and samples i) and y,!,, an unbiased YY Ke) ae x We conducted two preliminary case studies in which the robustness of the SCANN generator, derived from the LEAP net, was evaluated on a 5-bus system: The training power flow data were obtained with Pandapower [41], am open-source Python library for power system modeling and snalyss. The datasets comprised zero (x) and r.~1 power lin fares with a predefined probability p and samples drawn from a Bernoulli distribution, Le. ~ 0 implies zero line disconnection while to be disconnected once. "The firs case ing the power flow under n~ contingency based on zero and one power line disconnections proved as 122 Auorzed cans use vedo Lula Univers of Technolgy. Downloaded on Apt 90 202 a 1708-17 UTC fram IEEE Xplore Resbichons acyNovonal SCAN architecture where the latent space KS presesined on mand n — 1 power flow da for piven ged topology. The diseriminuor etaluites generated ~ (K-41) power flow against afew tre dats samples for > and classifies them as real oc ake accordingly. The generated samples ‘lasted as real” are fd Hock to the peerator fr subsequent lem true datasets, whereas the second case predicted based on true and n~ 1 datasets as well as on new n ~ 2 data generated by the network. Fig. 3 depicts preliminary results wherein the ls-loss of power line current is plotted as a function of p. According to the figure, the exploitation of newly generated mm — 2 data significantly improves the n — 3 power flow prediction accuracy while retaining its rapid convergence at a low 2p. The generator outperforms ResNST in all cases, and we expect the ongoing SCANN development to further improve the prediction accuracy and convergence. ‘Our ongoing research efforts as part of the SCANN de- velopment include the formulation of neural networks for the discriminator and selection of an appropriate kernel K’ in Eq, (1). In addition, SCANN will be complemented by a rnovel global resilience characterization tool called Resilience- Chaos plots (RCPS), which is under development to provide practical insights into worst contingencies for utilities. The abstract concept of RCPs is described in the following section IV, RESILIENCE-CHtaos PLOTS The RCP is a new type of resilience characterization tool introduced here as a quick way to analyze contingency pre- dictions. RCPs are composed of the contingency order (k) on the z-axis, which represents the amount of chaos imposed on the system (i¢., loss of any single or multiple components ddue (0 a adverse event), and a resilience performance index Rk) on the y-axis (See Fig. 4), Here we clarify that the term “chaos” does not refer to its mathematical definition; instead, it connotes the unpredictability of adverse events (e.g. where the next hurricane will land, what equipment will be damage by lightning, or when the next successful eyberattack will occur). Common resilience performance indices include % under-voltage busses [42], % load served [43], or system adaptive capacity [44]. The resilience metric is evaluated for un predictions, The oneal arhtetare fs been simple for ilaxration purpose, 2axio r + 1x0? eC) (2) Pressed on trae and — 1 datets Hi fe ql see 0 Wa a () Presrined on ue rand m= 1 dasets as wel fs ely genorated m — 2 data Fig. 3. Proiminary supe gencralization resus (n~3 power ow peition) ‘obianed with the SCANN generator underdevelopment the worst ease in a given set of m — k contingencies, and it is plotted as a function of F until an unacceptable level of performance is reached Fig. 4 depicts an RCP of notional data to exemplify the ‘characterization of a system under transformer overloading, where operating transformers at an overloaded capacity for 123 vores tensed se lmtedto:Lulea Univers of Tecnology, Dawnoaded on Apt 902024 at 170817 UTC om IEEE Xplore Rastictons ace15, 30, 60, 90, 180 minutes may correspond to permissible levels of 150%, 140%, 130%, 120%, and 110% beyond the normal rated capacity, respectively. Ifthe transformers operate at 150% capacity, for instance, the system can withstand all n— 8 contingencies. System operators, however, have only 15 minutes (© alleviate the extra electrical stress before the transformers incur damage or begin (0 trip offline. The use of RCPS could thereby help utilities quickly determine the cost-effective upgrade along with computationally inexpensive contingency prediction tools such as SCAN, ‘The minimum normalcy in Fig. 4 is an important concept that defines the tolerance a grid may deviate from its optimal operating state before exhibiting unacceptable performance [45]. The minimum normalcy may vary from utility to utility and be established by various operational regulations (such as NERC standards for example) or internal standards used by the utility. The point at which the grid performance imtersects the minimal normalcy line is called the resilience break point (RBP), and it represents the amount of chaos a grid can tolerate before exhibiting unacceptable behavior. RCPS can function as a reward mechanism in reinforced learning to suggest remedial actions or new system upgrades in ‘ovo ways: 1) maximize the area above the minimal normalcy for 2) extend the resilience break point out to a higher &. Values along the y-axis for the undisturbed system (i.e. k = 0) should be the highest for the system; hence, the ideal curve above ‘minimal normalcy should be a constant line up 10 a specific kk for which the system was designed. Although a nonlinear decay such as in Fig. 4 features some tolerable performance degradation, retaining the system performance as optimal as possible (nearly constant) until reaching the RBP is typically ‘more desirable especially when R(k) = % load served. ‘The difference between the ideal and actual /2(f) can be deseribed <= Omins 1S min 20 30min 60 min 50 min 2180 min Minimal Resilence fess, contingencies Resilient 20 | forall mt Performance index (e.g % load served) ig. 4. Resiience-Chaos Plt of notional data for six diferent transformer overloading time schemes in terms of the fill factor (ff) as follows: Ak SS RG) + RR +1) ff 0% cave 2 where knpp is the contingency order before the RBP and Ak=1. ‘The level of importance between increasing ff and ken must be assigned on a case by case basis. Ideal system upgrades will improve both values, but utilities may place a higher premium on ff for areas with critical loads or high value contracts and customers where service degradation is not acceptable, Conversely. if power outages occur frequently within an area, upgrades that increase yp may become more attractive option, V. CONCLUDING REMARKS The brief survey of previous studies on ML techniques and power grid analyses corroborates the remarkable potential of ML for power flow predictions under grid contingencies. Techniques such as super-generalizarion have particularly su ceeded in accurately predicting n — 2 contingencies based on n — 1 training data. In an effort to cope with limited training data for supervised learning, INL and FSU have teamed up to develop SCANN, a zero-shot and deep learing- ‘based contingency analysis tool exploiting CGAN and LEAP. net for unsupervised learning. Furthermore, we introduced the preliminary concept of RCPS as a novel resilience met ‘complementing SCANN which may prove to be a practical aid when SCANN succeeds in the utility industry. We expect to publish in the near future a detailed technical descri of SCANN and RCPs along with meaningful results and discussions on their advantages, drawbacks, and practicality REFERENCES. 11] E.C.R. Council “Th economic impacts ofthe avgust 2003 Mackout” Wisvhirgiom, BC, 2008, PK: Fazzini and DiChestopher, “Am slamingly simple eyberatck hit ‘crc systems serving la i i ake, Power Never Won Ge ‘CNBC, 2019. |Online) Available: hups:/ww.cnbe.com201 90503 (Bos sitek-causd-nterrupions-n-power-ste-operaios- doe hn 1B. Soberak, “denial of service” lack cated rd cyber dlmipion: Dee" EAE News, 2019. [Onlin] ‘abe ipsum cenews nostri 1002847518! {Eat=A% 20 cen 2neyhers Open Mo 20Departnen 220 nerey420o ical teat-Denia 2D Dae ie OC% 2ooc vit atcomputers% 200% 20operate7Onormaly HA Sit, J. Jatin, and 0. Alssg, ‘De power low revista” Ieee Transactions on Power Systoms, yl. 24, no. 3, pp. 1290-1300, 2008, [5] MK Enns J. Quads, and B.Sacket, “Fast linear contingency analysis” IEEE Tromacions on Power Apparai and Systems, Wo. 8 Pp. 785-791, 1982 [ot ¥-Brandwajn,“Elicint bounding method for tne contingency aal- sig TEBE Trnazclons on Power Systems, sol 3, 0. eps BAS, 198%, Y, Donde, V. Liper, B. Lescure, A. Ping, C. Yang, and J, Mera “Severe mulliplecomingeney serening in lee pow sms" THEE Transactions on Power Systems wl. 3, mo. 2, pp 406-417, 2008, [8] CoM. Davi and TJ Overbye,“Mullipe element contingency srcen- ing!" IEEE Transactions on Power Systems, vo. 28, no. 3. pp. 4 1301, 2010. 124 vores eensed se lmtedto:Lulea Univers of Tecnology, Dawnoaded on Apr 902024 at 170817 UTC om IEEE Xplore Restrictions ace[91 M.A, Eppstsin and PD. Hines, °A “random chemistry” agoritn for identifying collections of mulipe eomingeaces tht inate cascading Faure” IEEE Transacons on Power Systems, ol 27, 0.3, pp. 1688- 108, 2012. P. Kaplunoich and K. Turtsys, “Fast and reliable sowening of w-2 coatingencis” IEEE Transactions on Power Systoms, vo. 3, m0. 6, pp 218-1252, 2016 Y' Mansour, Chang J. Tay, E. Vashi, B. Corns and M. El Sarkawi, "Large scale dynamic sccuity screning and ranking using oul noiwotks" IEEE Transactions on Power Systems, wl. 12; 96.2, pp- 981-960, 197, [12) L: Webenkel, "Machine leaming approaches fo powersysem security assessment” EEE Expert. vo. 12.0. 5 pp. 4072 197. 1131 5, Reface, M. Motandes, and H. Magirabi, “Radial si function network for contingency analyse of balk power syste” ‘actions on Power Systems, vl. 14 0.2, pp. TIZTIR, 1989 (U4) Co'Stwastava”S. Singh, and J. Sharma. "A hybrid’ neural setwork model for fist votage contingency serening and ranking” dernavional Journal of Electrical Power & Energy Systems, ol 22, 0.1 pp 35-2, 200, “TS. Sis and Cu, "Contingency sesening for sod state secu ‘alysis by sng Mandarina networks" IEEE Transactions ‘in Power Stems, vO. 18, 1, p 421-42, 200. KK Nian, ©. Arora, and 8. Surana, "Power system security ewakation ‘sing aon: feature clacton wing divergence” Elecine Power Sytem Resear, vol. 69, no. 2-3, ps 161-167, 2004 KS. Soarup and G, Sodhakar, “Nevral civork, approach to con- lingeney screening and ranking in power systems.” Newroconpuring. Yo 70, no. 13, pp 105-118, 2006, D, Semichos and N. Avours, “Steady state contingency analysis fof eleuical networks using machine Ieaing techniques” in JFIP Insernaional ‘Conference on Artifical Ineligence Applications ad Inosanins. ‘Spring. 2006 pp. 281-289 Singh an Srivastava, “Tine ow contingency seletion nd ranking sing cascade curl notwork” Newocompuaing, vol. 70, 90. 16-18, Pp. 26-268), 2007 KS. Swaryp “Anifcial neural network using pattern recognition for security assent a anal” Neuracompua, YO. Th 0.6, 583-998, 2008 (C Rudin, D. Walu, RN. Anderson, A. Boulanger, A Salle. Aouiss, M. Chow, H. Duta P'N. Gross, By Huang. S.orome era, "Machine reaming fr the new york ety power grid FEEE ransacons on aera ‘analis and mache inelgence, vol. 34, no. 2 pp. 328-345, 2011 [22] KC Verma and K. Nii, “Supervised leaming apposeh 1 line contingency screcning and ranking im power sytem" dnterarionl Sournal of blerrical Pier & Energy Stems, ol. 38 0. 1, pp. 97 Tos, 2012 [23] B. Donwoi, L Guyon, M. Schoonauer, P Pancaici, and A. Maro “Inducing machine Ieaming for ower sytem operation sper arXiv preprine aXXs" 170809597, 2017 B, Donact, 1. Guyon, M. Schosnaner, A. Maro, and P,Pancatic, “Fast power system security analysis wilh guided dropout” aXiv preprint i180 09870, 2018 7 amtipating comtingengcs in power rid using fas neural net ‘crwshing, ia 2018 International Join Conference on Neural Nerworks (HCNN). WEEE. 2018 pp. 18 (C-Ki, K. Kim, P. Baaprakash and M. Antescu, “Graph convolutional cual nctwork for optimal load shedding underline contingency” io 2019 IEEE Power & Energy Society General Mecing(PESCM). IEEE, 2019, pp. 1. Bi, Donon, B, Donnot, L Gaon, and A. Marot, “Graph neural solver fe pomer systems” in 2019 fcmavonal Joint Conference om Neural ‘Nenworks (ICN) "IEEE, 2019, pp. 1-8 BB. Donon B. Dont, L Gayoa, 7 Liu, A. Mar, P. Panes, and MM. Sehoenauer, “Leap nts for sym sentificaion and application to power systems.” Newrcompung, 2020 ©. Brosindy, D. Westermann, and R. Krebs “Recent and prospective “evelopments in power system conuol centers Adaping the digital vin lcchnology for application in powcr sper como centr in 2018 IEEE International Exergy Conference (ENERGYCON). IEEE 2018, pp. 1 GE Cooper and E, Herskovits, “A bayesian metal forthe induction ‘of probit networks from da” Machine Tearing, wl. 8, m0. pp. 309-347, 192, v0} un ust Lis} 07 st 91 0) pu a) bs) es} ea ro po 130) bu i bal Bs} ba i bs 191 [so ta isa us} 128 Autborzedlicensed use lmted to: Lulea University of Technology. Downloaded on pd 30-2024 at 1708:17 UTC from IEEE Xplre. Resbcions aly. M. Ricdmillr and Ht Haan, ‘backpropagation leaning: The rpop ager,” in IEEE inernsional conference on neural networks 1093, pp. 386-58 KH, X. Zhang. . Renan. Sin, "Deep rsa Karning For image ‘weoghition,”in Proceedings ofthe TEEE conference om computer ison and pasern recogtton, 2016, pp. T1-T7. Otice oF Hecusiy, Nov “American Energy Resllenee Model, 2019, [Onlin]. Available: haps cncry goWsitesproaHls/2019 (O7/feSNAERM_Repwrt_pubic_verson_072219_508.p 1. Toons, —"Resilentdisisbution systems pontoio ‘verview 2018, DOK GMI Peer "Review. [Online ‘Avalable: hpsstewscencrgy goulsiesproales2O1NTSSresial_ istitution systems. overview proj presentations pr rcarc Power Ressarch Insti, Arica! Iuelignce: Conceps for Blecrc Power, 217 "Machine Leamting Telwiques Using Synchrophasor Daa: Event Dareton and Ideeiicarion, 2018, LSeesne, Karangeos, and L. Webenkel, “Recent developments in machine Icing for energy nylons reliability management” r= ceedings of the HELE, 2000 MMi and 8. Osindero, “Conditional generative adversarial nets ‘arXiv preprne arXiv 14711784, 2014 4 Huang, A. Grtion, K. Borgwand, B. Schihopf and A.J. Snot, cometing samp selection bis by unlabeled data” in Advances On neural information processing systems. 2007, pp. 601-08, C.Fsichan, 5. 1. Hyland, and G. Ratscy "Rea-valued (media time series gencrajon with reuren!condional gan.” aPX preprint ‘ary 77002655, 217, LE Thurer, A. Scheider Schafer J. Menke, J. Dolion, B Mee, 'S’Meinecke, and M. Braun, Spandapower— am opensource python tol {or comenicnt modeling, analysis snd optimization of lectric ower ‘S)Mhn” PEEE Transactions om Power Stems, 0. 33,00. 510- (521, Nov 2018, AA. Keasinski, “Quanitative model and mets of electrical gris’ ‘eallenc evalated a power dsinbution evel.” Eergies, v9. 90.2 p98. 2016. P Ciclo, T Swat, B. Vagensith, Reger, J. Gentle, TMeJunkin, and FColllaSoncher, “Pectrical grid resilience framework with Ueenainty” Eleeine Power Syiems Kescarch, ol. 180. p. 106801 ino, BB. Vangensmith, T. Melunkin, K. Vodrs, J. Reeves, J. Waymeat, Bosre, C. Ricgsr, and J. Case, “A inept appro to improving power grid olay” Merging of protic sk assent With Feslence meuics.” in 2018 Resilience Week (RWS). IEEE, 2018, pp 0-146 CG. Ringer, “Resilint contol spsoms proticel motes bass for ‘etning mission impact.” in 2014 7eh Imeradonal Sympostam on Resins Conl Systems (ISRCS).. UKE, 2014, pp. 1-10.