Table Of Content

1 Unit Commitment on the Cloud Mushfiqur R. Sarker, Member, IEEE, and Jianhui Wang, Senior Member, IEEE Abstract—TheadventofHighPerformanceComputing(HPC) the performance gain of the UC problem on the cloud. has provided the computational capacity required for power The objective of this work is to introduce the benefits and system operators (SO) to obtain solutions in the least time challenges of outsourcing the UC application to the cloud. to highly-complex applications, i.e., Unit Commitment (UC). TheUCissolvedusinga large-scalepowersystemtoanalyze The UC problem, which attempts to schedule the least-cost combination of generating units to meet the load, is increasing the computational performance under different categories of incomplexityandproblemsizeduetodeploymentsofrenewable cloud-based instances with comparisons to a traditional in- 7 resources and smart grid technologies. The current approach to house HPC infrastructure. 1 solvingtheUCproblemconsistsofin-houseHPCinfrastructures, 0 whichexperienceissuesatscale,anddemandshighmaintenance II. BENEFITSAND CHALLENGESWITH UCON THECLOUD 2 and capital expenditures. On the other hand, cloud computing The powerful computationalcapacity, rapid scalability, and n isanidealsubstituteduetoitspowerfulcomputationalcapacity, a rapid scalability, and high cost-effectiveness. In this work, the high cost effectiveness are the three major benefits to cloud J benefitsand challenges of outsourcing theUC application to the computing that SOs can exploit. Specifically, for example, 8 cloud are explored. A quantitativeanalysis of thecomputational Amazon’s cloud product, EC2, provides up to 128 virtual 1 performance gain is explored for a large-scale UC problem CPUsand1952GBofRAMpercloudinstance,whichcanbe solved on the cloud and compared to traditional in-house HPC provisionedin a shortperiodof time. Theability to outsource ] infrastructure. The results show substantial reduction in solve C time when outsourced to the cloud. simulationstotheclouddecreasestherequirementforin-house D HPC infrastructures, thus providing potential savings in both IndexTerms—Cloudcomputing,high-performancecomputing, cost and computational overhead. . unit commitment, power system applications s Although there are major benefits of the cloud, there are c [ I. INTRODUCTION significant challenges related to the cybersecurity aspects The recent paradigm shift in the power system industry SOs must consider. Three overarching cybersecurity chal- 1 lenges must be addressed for operation-based applications: v is to take advantage of High Performance Computing (HPC) (1) infrastructure security, (2) data confidentiality, and (3) 6 infrastructures to solve operation-based applications, such as 8 Unit Commitment (UC) [1]–[3]. UC is a highly-complex time criticality. Within (1) infrastructure security, the local 8 and cloud infrastructures, and the communication between mixed integer linear program (MILP) which determines the 3 them must be secured from potential insider and outsider commitment status of participating supply providers (e.g., 0 cyberattackers with intentions of passive (i.e., eavesdrop on . conventional units, renewable resources, and even demand- 2 thedatastreams)oractive(i.e.,maliciouslyperformfalsedata side resources) to meet the net demand at each operating 0 injections)manipulations.Furthermore,(2)dataconfidentiality 7 bus in the system, while adhering to supply provider and is crucial for operation-based applications, such as UC. The 1 transmissionconstraints[3],[4].TheneedforHPCarisesfrom UC problem includes generation- and network-specific data, : the increasing amount of renewable resources and smart grid v whichmustremainconfidentialandthusmechanismsmustbe i technologies, which has enlarged the computationalcomplex- X establishedtosecuredatafromcyberattacks.Lastly,thesecure ity and problem size of the UC application. Recent devel- r opments of solving UC with HPC infrastructure in the open outsourcingofUConthe cloud,mustconformtothe (3)time a criticalityrequirementssetforthbySOs.Theirwillbeinherent literature (e.g., in [2], [3]) and also in real-life operations tradeoff between enacting enhanced cybersecurity measures (e.g., by ISO-NE [1]) discuss two major benefits: 1) reduced andthesolvetimeforUConthecloud.Thechallengeremains computationtime,and2)inclusionofmechanismstoconsider in finding solution mechanisms that ensure high security and uncertainty of renewable resources. Traditionally, however, low computational overhead. HPC infrastructure is hosted by SOs in local computing environments(e.g.,byISONewEngland[1]),wheretheflexi- III. UNITCOMMITMENT FORMULATION bilitytoenhancecomputationalcapacitybecomesabottleneck, The compact matrix formulation of the SCUC problem is because marginal performance increases require high capital shownin(1)-(6).Theobjectivefunction(1)minimizesthesum expenditures and maintenance. ofthecommitmentcostscTz (i.e.,start-up,no-load,andshut- On the other hand, cloud computing is emerging as a new downcosts)anddispatchcostsbTyovertheoperatinghorizon. paradigm shifting technology to solve complex power grid The binary variable z ∈ {0,1} is a vector of commitment applications. The outsourcing of UC to the cloud enables related decisions, including the ON/OFF and start-up/shut- tapping into its powerful computational capacity, rapid scala- downstatus of each generationunitwithin each time interval. bility,andhighcosteffectiveness.Pioneeringsystemoperators The continuous variable y is a vector of dispatch related (SO) have explored the implementation of power system decisions, including the generation output. Equations (2)-(6) planning applications on the cloud, such as by ISO New contain commitment and dispatch related constraints. For a England [5], [6]. However, limited works exist on analyzing detailed formulation, the interested reader is encouraged to 2 TABLEI 15 +14.45% COMPUTINGINFRASTRUCTURECHARACTERISTICS 10 +9.43% CPU RAM SSD IntelProcessor 1)ANLBlues 16 64 X XeonNehalem 5 +1.95% 2)c4.2xlarge 8 16 X XeonE5-2666v3 0 3)c4.4xlarge 16 30 X XeonE5-2666v3 -3.74% 4)c4.8xlarge 36 60 X XeonE5-2666v3 -5 c4.2xlarge c4.4xlarge c4.8xlarge m4.16xlarge 5)m4.16xlarge 64 256 X XeonE5-2686v4 Computing Infrastructure Fig.1. Performance gainorloss(%).Foreachinstance, thepercentchange wasdeterminedusingtheaveragecomputationtimeoveralltrialsonAmazon refer to [4]. EC2againsttheaveragecomputation timeonANLBlues. min cTz+bTy (1) z,y s.t. Fz≤f, (2) ings.However,m4.16xlarge,whichincludes64CPUsand256 GB RAM, is outperformed by c4.8xlarge by +5.02% with 36 Hy≤h, (3) CPUs and 60 GB RAM. The M4 family of instances are not Az+By≤g, (4) tuned primarily for computation-intensive applications, rather Iuy=d, (5) they are for balanced applications requiring compute, high z∈{0,1} (6) memory, and network resources. An analysis such as done in Fig. 1 must be performed by IV. SIMULATION RESULTS SOs to determine the performance gain unique to the cloud To analyze SCUC performance, the Illinois, USA power provider of their choice. With the reduction in solve time, system was used, which lies within the Midcontent Indepen- the SO may keep the UC problem as-is, change the time dent System Operator’s (MISO) region. This system consists intervals by increasing the time horizon (i.e, 24-hour to 36- of 210 generators, 1908 buses, and 2522 transmission lines. hour look-ahead) or decreasing the granularity (e.g., 1-hour The SCUC model consists of four piecewise linear cost to 30-minutes), or adding enhancements such as uncertainty segments to preserve linearity. The SCUC leads to a highly- management of renewable resources. complex MILP problem with 237,817 variables, of which V. CONCLUSIONS 55,440arebinary.Themodelisformulatedbasedon[4]using This objective of this work was to introduce the benefits GAMS 24.0.1[7] and solved using IBM’s CPLEX [8] solver. and challenges of outsourcing a highly-complex and crucial The optimality gap was set to 0.5%. application, Unit Commitment (UC), to the cloud. By using For computational comparison, Amazon EC2 cloud in- the cloud, the SO receives up to a 14.5% savings in the stances [9] and Argonne National Laboratory’s Blues HPC computation time. While their are significant benefits to the (ANLBlues) were deployedto solve the SCUC. AmazonEC2 cloud, the cybersecuritychallenges will hinder its widespread is a cloud platform that provides rapid scalability of compu- adoption, thus mechanisms need to be enacted to secure UC- tational resources [9]. The local and cloud infrastructures are related data and the infrastructures involved. summarizedinTableI.TheAmazonC4instancesareequipped with high-performance processors ideal for computationally REFERENCES intensive applications, whereas the M4 instances provide an [1] “Evaluation of robust unit commitment by iso overallbalanceofcomputing,memory,andnetworkresources. new england: 6 month update.” [Online]. Available: www.hpc4energy.org/incubator/evaluation-of-robust-unit-commitment A. Computational performance on the cloud [2] A. Papavasiliou and S. S. Oren, “A comparative study of stochastic Since cloud instances are shared resources, the computa- unit commitment and security-constrained unit commitment using high tionalresourceavailabilityatanygiventimemaybedifferent. performance computing,” in Proc. of the European control conference, 2013. Therefore,to obtainan averagesolvetime, MonteCarlo trials [3] A.Papavasiliou,S.S.Oren,andB.Rountree,“Applyinghighperformance were performed, where the UC was solved for 100 trials computing to transmission-constrained stochastic unit commitment for on both cloud and local ANLBlues instances. To analyze renewable energy integration,” IEEE Trans. on Power Systems, vol. 30, no.3,pp.1109–1120, May2015. the performance gain or loss, the average percent change [4] H.Pandzic, T.Qiu, andD.S.Kirschen, “Comparison ofstate-of-the-art was calculated by comparing each cloud instance against transmission constrained unit commitment formulations,” in 2013 IEEE the ANLBlues’s average. Fig. 1 shows the average percent PowerEnergySociety GeneralMeeting, July2013,pp.1–5. [5] E.Litvinov,F.Ma,Q.Zhang,andX.Luo,“Cloud-basednext-generation performance gain or loss for each Amazon EC2 instance. itparadigm fortheoperations offuture power systems,”in2016Power From Fig. 1, compared to ANLBlues, c4.2xlarge performs SystemsComputation Conference (PSCC),June2016,pp.1–7. ata computationalperformancelossdueto 8CPUs compared [6] J. Goldis, A. Rudkevich, L. Omondi, and R. Tabors, “Use of cloud computing in power market simulations,” FERC Technical Conference to the 16 CPUs available within ANLBlues. However, as the on Increasing Real-Time and Day-Ahead Market Efficiency through number of available CPUs are increased, the C4 family of ImprovedSoftware, Report, 2014. instances (i.e., c4.4xlarge and c4.8xlarge) provide a positive [7] “Gams - a user’s guide.” [Online]. Available: www.gams.com/dd/docs/bigdocs/GAMSUsersGuide.pdf performance gain compared to ANLBlues. The increased [8] “Ibmilogcplexoptimization studiouser’smanual,”IBM,Report,2015. performance between c4.4xlarge to c4.8xlarge indicates that [9] “Amazon ec2 instances - documentation.” [Online]. Available: marginal increase in CPUs leads to computational time sav- docs.aws.amazon.com/AWSEC2/latest/UserGuide/Instances.html