ScienceDirect ScienceDirect

16 downloads 2694 Views 774KB Size Report
2016, IFAC (International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved. the effectiveness of the proposed approach. Keywords: ...
IFAC and CIGRE/CIRED Workshop on IFAC on Control ofCIGRE/CIRED TransmissionWorkshop and Distribution Smart Grids IFAC and and CIGRE/CIRED Workshop on IFAC and CIGRE/CIRED Workshop on Control Transmission and Smart Grids October 11-13, 2016. Prague, Czech Republic Control of of Transmission and Distribution Distribution Smartonline Grids Available Control of Transmission and Distribution Smart Grids at www.sciencedirect.com October October 11-13, 11-13, 2016. 2016. Prague, Prague, Czech Czech Republic Republic October 11-13, 2016. Prague, Czech Republic

ScienceDirect

IFAC-PapersOnLine 49-27 (2016) 445–450

Machine Learning Techniques for Power Machine Learning Techniques for  Machine Learning Techniques for Power Power System Security Assessment System Security Assessment  System Security Assessment

Nikita Nikita Nikita Nikita

V. V. V. V.

Tomin, Tomin, Tomin, Tomin,

Victor G. Kurbatsky, Denis N. Sidorov, Victor G. Kurbatsky, N. Sidorov, Victor G. Kurbatsky, Denis Alexey Zhukov Denis Victor G.V. Kurbatsky, Denis N. N. Sidorov, Sidorov, Alexey V. Zhukov Alexey V. Zhukov Alexey V. Zhukov Melentiev Energy Systems Institute, Irkutsk, 664033 Russia (e-mail: Melentiev Energy Energy Systems Institute, Irkutsk, 664033 Russia (e-mail: Melentiev Institute, [email protected]). Melentiev Energy Systems Systems Institute, Irkutsk, Irkutsk, 664033 664033 Russia Russia (e-mail: (e-mail: [email protected]). [email protected]). [email protected]). Abstract: Modern electricity grids continue to be vulnerable to large-scale blackouts. As all Abstract: Modern electricity grids continue to there be vulnerable to large-scale blackouts. As all Abstract: Modern electricity grids to be to blackouts. As states leading to large-scale blackouts are unique, is no algorithm to identify pre-emergency Abstract: Modern electricity grids continue continue to there be vulnerable vulnerable to large-scale large-scale blackouts. As all all states leading to large-scale large-scale blackouts are unique, unique, is no algorithm to identify pre-emergency states leading to blackouts are there is no algorithm to identify pre-emergency states. Moreover, numerical conventional methods are computationally expensive, which makes states leading to large-scale blackouts are unique, there is no algorithm to identify pre-emergency states. Moreover, numerical conventional methods are computationally expensive, which makes states. Moreover, conventional are which it difficult to use numerical for the on-line securitymethods assessment. Machine learningexpensive, techniques withmakes their states. Moreover, numerical conventional methods are computationally computationally expensive, which makes it difficult to use for the on-line security assessment. Machine learning techniques with their it difficult to use for the on-line security assessment. Machine learning techniques with their pattern recognition, learning capabilities and high speed of identifying the potential security it difficult to use for the on-line security assessment. Machine learning techniques with their pattern recognition, learning capabilities andThe high speedofof ofthis identifying the to potential security pattern recognition, learning capabilities and high speed identifying the potential security boundaries can offer an alternative approach. purpose paper is not suggest that one pattern recognition, learning capabilities and high speed of identifying the potential security boundaries can of offer an alternative alternative approach. The purposeassessment of this this paper paper is not notbeto tomore suggest that one one boundaries can offer an approach. The purpose of is suggest that particular kind machine learning technique for security would appropriate boundaries can of offer an alternative approach. The purposeassessment of this paper is notbetomore suggest that one particular kind machine learning technique for security would appropriate particular kind machine technique security assessment would more appropriate than others. Weof start fromlearning the premise that for almost every method may bebe useful within some particular kind of machine learning technique for security assessment would be more appropriate than others. We start start fromonthe the premise that almost every every method may may be useful useful approach within some some than others. We from that almost method be within restricted context. Based thispremise idea, we developed an automated multi-model for than others. We start from the premise that almost every method may be useful within some restricted context. Based on onThe thisproposed idea, we we method developed an automated automated multi-model approach for restricted context. Based this idea, developed an multi-model for on-line security assessment. allows us to automatically test approach the different restricted context. Based on this idea, we developed an automated multi-model approach for on-line security assessment. Thetoproposed proposed method allows us to and automatically test the different on-line security assessment. The allows us automatically test state-of-art techniques in order find bothmethod the best algorithm its top performance tuning on-line security assessment. Theto proposed method allows us to to and automatically test the the different different state-of-art techniques in order find both the best algorithm its top performance tuning state-of-art techniques in order to find both the best algorithm and its top performance tuning for particular analyzed power system. A case study using the IEEE RTC-96 system demonstrates state-of-art techniques in order to findAboth the best algorithm and its topsystem performance tuning for particular analyzed power system. case study using the IEEE RTC-96 demonstrates for analyzed system. A the particular effectiveness of the power proposed approach. for particular analyzed power system. A case case study study using using the the IEEE IEEE RTC-96 RTC-96 system system demonstrates demonstrates the effectiveness of the proposed approach. the effectiveness of approach. the effectiveness of the the proposed proposed approach. © 2016, IFAC (International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved. Keywords: smart grid, power system, security assessment, blackout, machine learning. Keywords: smart grid, power system, security blackout, machine learning. Keywords: smart smart grid, grid, power power system, system, security security assessment, assessment, Keywords: assessment, blackout, blackout, machine machine learning. learning. 1. INTRODUCTION blackouts. The decision making and onus is usually still 1. INTRODUCTION blackouts. The decision and onus is usually still 1. INTRODUCTION INTRODUCTION blackouts. The making and with the expertise of the making grid operators. 1. blackouts. The decision decision and onus onus is is usually usually still still with the the expertise expertise of the the making grid operators. operators. with of grid The security of a power system is related to the ability with the expertise of the grid operators. For the time being there is a wide spectrum of approaches The security of power system is related to the the despite ability For the time being there is a wide spectrum of approaches The security of aaa power power systemnormal is related related to ability of a security power system to continue operation For the time there spectrum approaches and tools for being the assessment of security. Theof obThe of system is to the despite ability the time there is is aa wide wide spectrum ofprimary approaches of a power system to continue normal operation of a power power casualties system to to to continue normal operation despite and tools for being the assessment assessment oftechniques, security. The primary obunplanned operating equipment, known as For and tools for the of security. The primary objective of security assessment then, is to meaof a system continue normal operation despite and tools for the assessment of security. The primary obunplanned casualties to operating equipment, known as unplanned casualties to of operating equipment, known as as jective of security assessment techniques, then, is to meacontingencies. A failure security equipment, can cause equipment jective of security assessment techniques, then, is to measure the vulnerability of the system to blackouts. Unfortuunplanned casualties to operating known jective of security assessment techniques, then, is to meacontingencies. A failure of security can cause equipment contingencies. A failure failureorof oflow security can and cause equipment the vulnerability of the system to blackouts. Unfortudamage, low frequency voltages, localized loss sure sure the vulnerability of to Unfortunately. evaluation of this measure is not within contingencies. A security can cause equipment sure thereal-time vulnerability of the the system system to blackouts. blackouts. Unfortudamage, low frequency or low voltages, and localized loss damage, low frequency or low voltages, and localized loss nately. real-time evaluation of this measure is not within of power to customers, but the most severe, spectacunately. real-time evaluation of this measure is not within the capabilities of current conventional technology. This is damage, low frequency or low voltages, and localized loss nately. real-timeofevaluation of this measure is not This within of power to customers, but the most severe, spectacuof power to customers, but the most severe, spectacuthe capabilities current conventional technology. is lar, costly and therefore most interesting security failures the capabilities of current conventional technology. This is due to the fact that numerical conventional methods are of power to customers, but the most severe, spectacuthe capabilities of current conventional technology. This is lar, costly and therefore most interesting security failures lar, costly and therefore therefore mostthe interesting security failures to the fact that numerical conventional methods are result in blackouts. During past ten security years events in due due to the fact that numerical conventional methods are computationally expensive, which makes it difficult to use lar, costly and most interesting failures due to the fact that numerical conventional methods are result in blackouts. During the past ten years events in result in blackouts. During Union the past past ten years events in computationally expensive, which makes it difficult to use North in America, European andten Asia have clearly computationally expensive, which it to for the on-line security assessment. Managing a modern result blackouts. During the years events in computationally expensive, which makes makes it difficult difficult to use use North America, European Union and Asia have clearly North America, European Union and Asia have clearly for the on-line security assessment. Managing a modern demonstrated an increasing likelihood of large blackouts. for the on-line security assessment. Managing a modern in real-time requires assessment. much more automatic monitoring North America, European Union and ofAsia have clearly grid for the on-line security Managing a modern demonstrated an increasing likelihood large blackouts. demonstrated an increasing likelihood of large blackouts. real-time much more automatic monitoring This indicates an that the security monitoring and control grid grid in in real-time requires much more monitoring and fast securityrequires assessment Machine learning demonstrated increasing likelihood of largeand blackouts. in real-time requires muchmeasures. more automatic automatic monitoring This indicates that the security monitoring control This indicates that theneed security monitoring and (2007), control grid and fast security assessment measures. Machine learning of power systems may to be monitoring improved IEEE and fast security assessment measures. Machine learning techniques with their pattern recognition, learning capaThis indicates that the security and control and fast security assessment measures. Machine learning of power systems may need to be improved IEEE (2007), of power systems may need to be improved IEEE (2007), techniques with their pattern recognition, learning capaWang et al.(2005), Syktyvkar (2010), Wehenkel (1995). techniques with their pattern recognition, learning capabilities and high speed of identifying the potential security of power systems may need to be improved IEEE (2007), techniques with their pattern recognition, learning capaWang et al.(2005), Syktyvkar (2010), Wehenkel (1995). Wang et al.(2005), Syktyvkar (2010), Wehenkel (1995). bilities and high speed of identifying the potential security bilities and high speed of identifying the potential security boundaries can potentially offer suchthe on-line solution. Wang et al.(2005), (2010), Wehenkel Most power plantsSyktyvkar and transmission lines are (1995). overseen bilities and high speed of identifying potential security boundaries can potentially offer such on-line solution. can offer such on-line Most power plants and transmission lines are overseen Most power plants plants and transmission transmission lines are are(SCADA) overseen boundaries by a supervisory control and data acquisition boundaries can potentially potentially such learning on-line solution. solution. Many researchers deem thatoffer machine (ML) methMost power and lines overseen by a supervisory control and data acquisition (SCADA) by a supervisory control and data acquisition (SCADA) Many researchers deem that machine learning (ML) methsystem. SCADA technology goes back 40 years. Much of Many researchers deem that machine learning (ML) methods such as artificial neural networks (ANNs), decision by a supervisory control and data acquisition (SCADA) Many researchers deem that machine learning (ML) methsystem. SCADA technology goes back 40 years. Much of system. SCADA technology goes back backand 40 does years.not Much of ods such as artificial neural networks (ANNs), decision it is tooSCADA slow fortechnology todays challenges sense ods such as artificial neural networks (ANNs), decision trees (DTs), deep learning models etc. are indeed able to system. goes 40 years. Much of ods such as artificial neural networks (ANNs), decision it is too slow for todays challenges and does not sense it iscontrol too slow slow for enough todays of challenges and does doesaround not sense sense (DTs), deep learning models etc. are indeed able to or is nearly the components the trees trees (DTs), deep learning models etc. are indeed able provide interesting security information in power systems, it too for todays challenges and not trees (DTs), deep learning models etc. are indeedsystems, able to to or control nearly enough of the components around the or control nearly enough of the components around the provide interesting security information in power grid. The result is that no single operator or utility can provide interesting in power (2010), security Wehenkelinformation (1995), Diao et al.systems, (2009), or control nearlyisenough ofsingle the components around the Syktyvkar provide interesting security information in power systems, grid. The result that no operator or utility can grid. Theorresult result is that that no single single operator operator orHowever, utility can can Diao et al. (2009), stabilize isolate a transmission failure. or at Syktyvkar Syktyvkar (2010), Wehenkel (1995), Diao (2009), Tomin et al.(2010), (2014).Wehenkel Actually, (1995), in their philosophy grid. The is no utility (2010), Wehenkel (1995), Diao et et al. al.machine (2009), stabilize or isolate aa phasor transmission failure. However, at stabilize or isolate isolate transmission failure. units However, at Syktyvkar Tomin et al. (2014). Actually, in their philosophy machine the transmission level, measurement (PMUs) Tomin et al. (2014). Actually, in their philosophy machine learining-based approaches are quite similar to existing stabilize or a transmission failure. However, at Tomin et al. (2014). Actually, in their philosophy machine the transmission level, phasor measurement units (PMUs) the transmission level, phasor phasor measurement units approaches are quite similar to existing havetransmission been introduced to improve grid reliability. One of learining-based learining-based approaches are similar to existing in power system security studies, where are the level, measurement units (PMUs) (PMUs) learining-based approaches are quite quite similar tolimits existing have been introduced to using improve grid reliability. One of practices have been introduced to improve grid reliability. One of practices in power system security studies, where limits are the issues of applying and the large amounts of PMU practices in power system security studies, where limits are derived from simulations, though in a manual fashion. But have been introduced to improve grid reliability. One of practices in power system security studies, where limitsBut are the issues of applying and using the large amounts of PMU the issues of applying and using the large amounts of PMU derived from simulations, though in a manual fashion. datasets are rapid decision making. Even if a lot of data derived from simulations, though in a manual fashion. But machine learning approaches are more systematic, easier the issues of applying and using the large amounts of PMU derived from simulations, though in a manual fashion. But datasets are rapid decision making. Even if a lot of data datasets are rapid decision making. Even if a lot of data machine learning approaches are more systematic, easier was available, the decision operatorsmaking. at different control centers machine approaches more systematic, easier handlelearning and master, in shortare more reliable and powerful. datasets are rapid Even control if a lot of data to learning approaches are more systematic, easier was available, the operators at different centers was available, theproper operators at different different control centers to handle and master, in short more reliable and powerful. did not take the actions in time control to prevent the machine to handle and master, in short more reliable and powerful. An important asset of machine learning methods lies in was available, the operators at centers to handle and master, in short more reliable and powerful. did not take the proper actions in time to prevent the did not not take take the the proper proper actions actions in in time time to to prevent prevent the the An important asset of machine learning methods lies in An important asset of machine learning methods lies in the explicit and logical representation they use for the indid An important asset of representation machine learning methods lies inin the explicit and logical they use for the the explicit and logical representation they use for the induced classification rules, which, together with simplicity, the explicit and logical representation they use for the in This work was supported by the Russian Scientific Foundation duced classification rules, which, together with simplicity, duced classification rules, together with  provide a unique explanatory capability, (2015). This work was supported the Scientific Foundation  duced classification rules, which, which, togetherSidorov with simplicity, simplicity, under and theRussian 2015 Endeavour ThisGrant work No. was 14-19-00054 supported by by the Russian Scientific Scholarship Foundation  provide aa unique explanatory capability, Sidorov (2015). ThisGrant work No. was 14-19-00054 supported by the Russian Scientific Scholarship Foundation provide unique explanatory capability, Sidorov under and the 2015 Endeavour provide a unique explanatory capability, Sidorov (2015). (2015). and Fellowship program. under Grant No. 14-19-00054 and the 2015 Endeavour Scholarship

under Grant No. 14-19-00054 and the 2015 Endeavour Scholarship and and Fellowship Fellowship program. program. and Fellowship program. Copyright © 2016, 2016 IFAC 445 Hosting by Elsevier Ltd. All rights reserved. 2405-8963 © IFAC (International Federation of Automatic Control) Copyright © 2016 IFAC 445 Copyright ©under 2016 responsibility IFAC 445Control. Peer review of International Federation of Automatic Copyright © 2016 IFAC 445 10.1016/j.ifacol.2016.10.773

2016 IFAC CTDSG 446 Nikita V. Tomin et al. / IFAC-PapersOnLine 49-27 (2016) 445–450 October 11-13, 2016. Prague, Czech Republic

The paper is organized as follows. Section 2 consists of two subsections. First subsection provides the problem statement in the field of modern security assessment problem. Second subsection presents the state-of-art in the field of on-line security assessment technologies. In section 3 a novel automated multi-model approach based on machine learning is developed for for online security assessment in power system. Section 4 presents the experimental results and discussion.

PRE-EMERGENCY CONTROL SECURE

NORMAL

Minimise the generation costs by optimising power flows according to the market situation

RESTORATIVE

ALARM

Resynchronization, load pickup, power supply restoration

Tradeoff of preventive vs corrective control

INSECURE

Preventive control

Emergency control (corrective)

EMERGENCY (NON-CORRECTABLE)

Protections

EMERGENCY (CORRECTABLE)

Overloads, undervoltages, underfrequency...

Cascading effects, blackouts

ASECURE

This work outlines some experience obtained at the Melentiev Energy Systems Institute, Russia in developing machine learning-based approaches for detecting potential dangerous states in power systems before they lead to major emergencies and blackouts. The proposed method allows us to automatically test the different state-of-art techniques in order to find both the best algorithm and its top performance tuning for particular analyzed power system. The calculations involved the different machine learningbased models, such as Multilayer Perceptron (MLP), Support Vector Machine (SVM), self-organized Kohonen network (SOM), Extreme Learning Machine (ELM), Random Forest (RF), Classification and Regression Tree (CART). A case study using the IEEE RTC-96 system demonstrates the effectiveness of the proposed approach. The suggested approach is implemented in the free software environment R intended for calculations with an open-source code.

Islands Load shedding

EMERGENCY CONTROL Control and/or protective actions

Foreseen or unforessen disturbances

Fig. 1. Operating states and transitions. Adapted from [Fink (1978)] 2.2 The state-of-art

2. BACKGROUND 2.1 Problem Statement Practical experience demonstrates that most blackouts begin with a large disturbance (a disturbance, which may or may not cause cascading failures), which leads to a slow deterioration of the system conditions, IEEE (2008), Muller et al. (2012), Lachs (2002). The system parameters may still remain within specified limits, but many of these parameters are on the boundary of stability; even a small additional disturbance can cause a simultaneous violation of several system parameters, and as a result, fast deterioration of the system state. If such conditions are identified as pre-emergency, preventive actions can be taken, and major events avoided. Unfortunately, in current competitive environment, such conditions may not be easily detected because different problems may simultaneously occur in different parts of a large network within different jurisdictions. The liberalisation process in power systems has created an additional interface which can adversely impact communication and coordination activities between operators on both sides. The past blackout events reveal that underlying causes are also partly linked to the liberalisation trends due to missing incentives to invest in reliable infrastructures. To monitor that the power system is within its limit, determined either online or offline, the primary measurement tools are SCADA systems and post processing by a state estimator, Morison (2004). The ENTSO-E network code on operational security requires each TSO to classify its system according to the system operating states ENTSO(2013). Figure 1 shows the different operating states of a power system as identified by Dy Liacco and adopted by authors of this paper. 446

In the last few years significant advances have been made in the field of on-line security assessment technologies, CIGRE (2007). In this report, there are reported 19 tools for dynamic security assessments in use, under testing and under development. A review of 15 of these state-of-theart tools shows a wide variety of implementations. The range of assessment capabilities includes determination of critical contingencies, transfer limits, and determination of remedial measures necessary to ensure security. The computational methods used for each type of security assessment is varied, and depends on the specific requirements, power system characteristics and, in some cases, the techniques available in the state-of-the-art tools used. However, conventional numerical techniques are usually time consuming and therefore are not always suitable for real-time applications. Also, these methods suffer from the problem of misclassification or/and false alarm. Misclassification arises when an active contingency is classified as critical. A great many studies show that the effective solution to this problem can be found on the basis of machine learning methods which normally include ANNs, DTs, deep learning models, etc. In the mid-eighties, it has already been demonstrated that machine learning is indeed an efficient and effective way to generate reliable and interpretable security rules from very large bodies of simulated examples, Wehenkel(1994), even for as complex systems as are real large-scale power systems. The extracted rules are found to express explicitly problem specific properties, similarly to human expertise, and hence may be easily appraised, criticized and eventually adopted by engineers in charge of security studies. The flexibility of the machine learning framework allows one to tailor the resulting information to analysis, sensitivity analysis and control applications.

2016 IFAC CTDSG October 11-13, 2016. Prague, Czech Republic Nikita V. Tomin et al. / IFAC-PapersOnLine 49-27 (2016) 445–450

447

Many power engineering deem that ML methods are indeed able to provide interesting security information for various physical problems and practical contexts. This is related to their capabilities of fast detection of the images, patterns (i.e. typical samples), learning/generalization and, which is important, high speed of identifying the instability boundaries. Moreover, by using synthetic information extracted by automatic learning, instead of analytical methods, much higher speed may be reached for real-time decision making. Further, in terms of data requirements, whereas analytical methods require a full description of the system model, the approximate models constructed via automatic learning may be tailored in order to exploit only the significant input parameters, Tomin et al. (2015).

In the paper, we examine the two most common resampling methods: 10-fold cross validation and and bootstrapping. The process of using half of the data for training and half for testing is actually a special case of crossvalidation, that is, two-fold cross-validation. In ten-fold cross-validation, the data is randomly split into 10 groups that contain the same number of cases (or approximately). The analysis is performed 10 times. Doing things this way allows for more accurate estimates of the reliability of the algorithm on the data as the analysis is performed several times with different sets, which permits the obtaining of distribution of measures of reliability. Another advantage is that 9/10th of the data is used in the training set at each iteration, and we can use all cases at both testing and training stages, Mayor(2015).

An important bias in comparative studies also may be due to the highly variable degree of expertise of the authors in the different methods they try to compare. Often, researchers compare their own favorite algorithm, for which they are presumably expert, with a set of competing methods, which they discover while doing the comparative study. For this reason, the compared algorithms often represent the state of the art only for the favorite method, and under such conditions highly biased conclusions may be reached.

The aim of bootstrapping is also to obtain a more precise image of the reliability of the model on the data. This is done in a different fashion. Instead of partitioning the data for training and testing, a random sample of n cases is selected N times from the original set with replacement (meaning that the same case can occur several times at each iteration), where N is the number of iterations and n is the number of cases. The analysis is performed on each of the samples independently, which gives mean and standard deviation for the estimates.

The very large diversity of ML models, which are available in the present time, makes it difficult to obtain honest comparisons, and this is the main reason why this kind of comparison has started only recently, in particular with power system security problems. Another aspect which may render the assessment of computational performances difficult, is related to manual tuning which is required with many heuristic methods and which may influence quite strongly the resulting performances.

3.2 Method

3. PROPOSED APPROACH The purpose of this work is not to suggest that one particular kind of method would be more appropriate than others. Rather, we start from the premise that almost every method may be useful within some restricted context, and summarize the respective strengths and limitations of the various methods so as to highlight their complementary possibilities. We endorse a philosophy of the no free lunch theorem, Wolpert (1996), which states, that given no prior knowledge of prediction problem, no single method can be said to be better than any other.

The paper proposes a novel automated multi-model approach based on machine learning for online security assessment in power system. Specifically, multiple ML models are first trained off-line using the bootstrap or crossvalidation (Fig. 2). For each candidate tuning parameter combination, a ML model is fit to each resampled data set and is used to predict the corresponding held out samples (Fig. 3). The resampling performance is estimated by aggregating the results of each hold-out sample set. Resampling methods try to inject variation in the system to approximate the models performance on future samples. These performance estimates are used to evaluate which combination(s) of the tuning parameters are appropriate. Once the final tuning values are assigned, the final model is refit using the entire training set. The optimal model from each ML technique is selected to be the candidate model with the largest accuracy or the lowest misclassification cost.

3.1 Resampling methods for evaluating classification accuracy of security assessment models It is common in ML that a portion of a data set is used to test the performance of the trained classifier. Security assessment learning-based models of power systems often have been developed without being tested. The apparent classification accuracy of such models can be optimistically biased and misleading. Data resampling methods exist that yield a more realistic estimate of model classification accuracy. Because these resampling procedures require no new data, they are relatively inexpensive and we can use these techniques to assess a model’s accuracy (under model conditions) and then decide if the model is worthy of field validation. 447

Fig. 2. The basic method of the proposed idea For the online applications, the final ”best” ML model is used to be the candidate technique with the best performance. As result, the system information is periodically checked and updated in order to account for changing

2016 IFAC CTDSG 448 Nikita V. Tomin et al. / IFAC-PapersOnLine 49-27 (2016) 445–450 October 11-13, 2016. Prague, Czech Republic

4.1 Performance Measures for Classification MATLAB

Quasi-dynamic simulation program

In current paper we need to use proper performance measurement metrics both for regression and classification problems. In case of classification task we use precision, recall and F-score for each class defined by following equations. All these metrics may be obtained from confusion matrix. Precision represents as the number of examples correctly classified as class divided by the number of all the examples labeled by the classifier as class c. Recall is the number of examples correctly classified as class c divided by the number of all the examples of class c in the data. F-score is a harmonic mean of the above.

Data generation

PSAT

Voltage, loads, power flow etc.

Data collection

New dataset

Feature attributes

Feature attributes

JAVA objects Training and tuning (replacing methods)

R environment Caret package

ML1 model (i) . ML1 model (k)

ML2 model (i) . ML2 model (k)

MLn model (i) . MLn model (k)

ML1 Model (optimal)

ML2 model (optimal)

MLn model (oprimal)

Performance estimator

(the largest accuracy, the lowest misclassification cost Final ML model

P recisionc =

Security index

Fig. 3. A general scheme of an automated ML-based technique for online power system security assessment system states as accurately as possible so that the offline trained alarm model may continue to perform well on the new system states. The ”best” ML model also can be updated by including new system conditions and, if required, the tuning parameters and even type of a ”best” model can changed after re-checking procedure. The final ML-based model is used on-line to classify the system operating state and, if required, to produce an alarm. An open-source environment R with caret package is used as a computing environment for proposed ML models design and testing. Operating conditions are all generated using the Power System Analysis Toolbox (PSAT) (Fig.3). PSAT represents a software environment with an open code for operation on the platform Matlab or GNU/Octave. It was decided to use the JAVA object as an interaction mechanism, which were called R-interface agent. This agent connects to specially created server based on Rserve. 4. A CASE STUDY

Recallc =

448

(1)

T rueP ositivesc T otalExamplesOf Classc

F score = 2 ∗

(2)

P recision ∗ Recall P recision + Recall

(3)

The F-score can be interpreted as a weighted average of the precision and recall, where an F-score reaches its best value at 1 and worst score at 0. 4.2 Security Index The class labels formation module is designed to build the training sample data using ”secure” and insecure” labels. In the case of supervised learning, machine learning model is given a set of pre-classified data labels and tries to discover a rule allowing us to mimic as closely as possible the observed classification. In this paper, the term security index (SI) indicates the system security level for a given system operating condition and a specified contingency. The SI is defined by calculating the line overload index (LOI) and voltage deviation index (VDI) as given by following expression, respectively:

LOIkm

The IEEE RTS-96 power system has been considered to test the proposed multi-model approach for voltage security monitoring and assessment. The modified system has 53 buses and dynamic elements to represent generators and loads.The proposed test system has structural features of many real power systems, which may face loss of stability. For example, interconnected power systems in Denmark, Sweden and some regions in Russia have similar structures with large generation in one part of the system and large load centers in another. Operating conditions are all generated by offline simulations using PSAT. For each operating scenario considered, N-1 contingency case (line/generator/transformer outages) is simulated and load flow solution by Fast Decoupled Load Flow method is obtained. Each operating condition is termed as a pattern. Each pattern is characterized by a number of attributes like load level, bus voltages, power generation, forming the components of a vector called pattern vector security assessment.

T rueP ositivesc T otalP redictedAsClassc

  Skm − Slim · 100, = Skm 0,

 min   |Uk | − |Uk | · 100,   |Ukmin |  V DIk = 0,  max    |Uk | − |Uk | · 100,  max |Uk | SI =

w1 ·

 nL

i=1

if Skm > Slim ;

(4)

if Skm < Slim .

if |Uk | < |Ukmin |; if |Ukmin | ≤ |Uk | ≤ |Ukm |.

if |Uk | > |Ukmax |;

LOIi + w2 · nL + nB

(5)  nB

i=1

V DIi

,

(6)

where Skm and Slim represents the MVA flow and MVA limit of branch k-m, |Ukmin |, |Ukmax | and |Uk | are the minimum voltage limit, maximum voltage limit and bus voltage magnitude of k-bus respectively; w1 and w2 are the weighting factors of system security; nL and nB represent the number of lines and buses respectively.

2016 IFAC CTDSG October 11-13, 2016. Prague, Czech Republic Nikita V. Tomin et al. / IFAC-PapersOnLine 49-27 (2016) 445–450

449

Evaluating the SI as given by , each pattern is labeled as belonging to one of the four classes as shown in Table I.

Cost 0.25 0.5

Table 1. Security class labels

1 2 0

Security Index SI = 0 0 < SI ≤ 5% 5% < SI ≤ 15% SI > 15%

Class Category/Power State Normal state Alarm state Emergency (correctable) state Emergency (non-correctable) state

In this paper, SI computation is used only as an aid for the formation of a class labels for a supervised ML model. We believe that the trained adaptive model can provide the better solution to correctly detect of alarm states based SI compared direct computation of this index. This is due to the fact that any algorithmic approach may require adjustment in the event of the new conditions, for example, changes in the system topology. A learning model is capable to respond to such changes adaptively. In addition, the real-time computation of SI for complex power system may take longer than the response of a trained ML model. 4.3 Training and Performance Machine learning models have been built for classifying the power system states, for various candidate attributes and four different security classifications. The models were trained on 6877 samples dataset and tested on 1715 samples (Fig. 4). Namely, the following state-of-art classification techniques were tested: Multilayer Perceptron, Support Vector Machine, self-organized Kohonen network, Extreme Learning Machine, Random Forest, Classification and Regression Tree. The models parameters were obtained using 2 repeating 10-fold cross-validation. As previously mentioned, the optimal model from each technique is selected to be the candidate model with the largest accuracy. If more than one tuning parameter is optimal then the function will try to choose the combination that corresponds to the least complex model. For example, for the SVM technique, scale was estimated to be 10, degree = 3 and cost C = 0.25 appears to be optimal (Fig. 5).

Accuracy (Repeated Cross-Validation)

degree: 1

4 2

4

6

8

10

degree: 2

degree: 3

0.99

0.98

0.97

0.96

0

2

4

6

8

10

0

2

4

6

8

10

Scale

Fig. 5. Density plots of the 10-fold cross-validation estimates of accuracy for the optimal SVM model. equal to 4. ELM model was used with number of neurons in hidden layer which is equal to 9 and sine activation function. As for supervised SOM model, the parameters were selected as xdim=5, ydim=6 and xweight=0.8. Table II shows comparison of accuracy achieved by the classification learning techniques. From Table II, the comparison indicates that the best performance on the test set using the final ML model (Random Forest), the accuracy of which is 99.89%. In other words, there are only 0.11% cases misclassified using this ML model in detecting dangerous states in the IEEE RTC-96 test system. By periodically including new and unknown system states into the database, MLs are updated to learn more useful information for improving robustness and the classification accuracy can be effectively increased. The proposed semi-automated method for online security assessment should work well for all kinds of unforeseen operating conditions no matter how the critical system parameters are distributed. Table 2. Performance measures by the 10-fold cross-validations

1.05

Models 1

Multilayer Perceptron

Load secondary voltages

voltage, p.u.

0.95

0.9

0.85

Load primary voltages

0.8

0.75

0.7 0

500

1000

1500

2000

2500

3000

3500

Performance Measures precision recall f-score 0.8912 0.9297 0.8903

Support Vector Machine

0.9847

0.9800

0.9821

Kohonen SOM

0.9969

0.9971

0.9970

Extreme Learning Machine

0.9437

0.9057

0.9222

Random Forest

0.9989

0.9992

0.9990

CART

0.9964

0.9969

0.9966

4000

time

Fig. 4. Voltage collapse simulation for IEEE RTS-96 power system Besides that MLP was used with number of neurons in hidden layer which is equal to 9. RF was used with mtry = 11 and 150 trees. CART was used with max tree depth is 449

5. CONCLUSIONS During the past ten years events in North America, European Union and Asia have clearly demonstrated an increasing likelihood of large blackouts. This indicates that the security monitoring and control of power systems may

2016 IFAC CTDSG 450 Nikita V. Tomin et al. / IFAC-PapersOnLine 49-27 (2016) 445–450 October 11-13, 2016. Prague, Czech Republic

need to be improved. This paper presents a novel semiautomated method for on-line security assessment using machine learning techniques. Multiple MLs, such as ANNs, SVM, decision trees etc., are first trained offline using the resampling cross-validation method. Resampling the training samples allows us to know when we are making poor choices for the values of ML tuning parameters. The best model from the ML techniques is selected based on its performance. For the on-line applications, the final the best of the best ML is used as the candidate technique with the best performance. If required, the final ML checked and updated in order to account for new changing system states as accurately as possible. The results showed that the proposed approach can identify potential dangerous states with high accuracy and, if required, the final ML model can produce an alarm for triggering emergency and protection systems. REFERENCES CIGRE Working Group C4.601 (2007), Review of on-line dynamic security assessment tools and techniques. Diao R. et al. (2009). Decision tree-based online voltage security assessment using PMU measurements, IEEE Trans. Power Syst., Vol. 24, No.2, pp. 832-839. Fink L.H., Carlsen K. (1978) Operating under stress and strain, IEEE Spectrum 15, No. 3, 4853. (4, 389) ENTSO-E. Network Code on Operational Security. 2013 24 September 2013; Available from: http://networkcodes.entsoe.eu/operationalcodes/operational-security/. IEEE PES PSDP Task Force on Blackout experience, mitigation, and role of new technologies, blackout experiences and lessons (2007), Best practices for system dynamic performance, and the role of new technologies. IEEE Special Publication 07TP190. IEEE PES CAMS Task Force on Understanding, Prediction, Mitigation and Restoration of Cascading Failures Initial Review Of Methods For Cascading Failure Analysis In Electric Power Transmission Systems, In Proc. IEEE PES General Meeting, Pittsburgh, PA USA July 2008 1. Lachs, W.R. (2002). Controlling grid integrity after power system emergencies, Power Systems, IEEE Transactions on, vol.17, no.2, pp.445-450 Mayor, E. (2015) Learning Predictive Analytics with R, Packt Publishing, 332 p. Methods and models for power system reliability studies, Syktyvkar (2010): Komi Scientific Center of Ural Branch of RAS, 292 p. (in Russian). Morison, K., Wang, L., and Kundur, P. (2004) Power system security assessment. IEEE Power and Energy Magazine, 2(5): p. 30-39. Mller S.C., Kubis A., Brato S., Hger U., Rehtanz C., Gtze J. (2012) New Applications for Wide-Area Monitoring, Protection and Control, In 2012 IEEE Proc. of the 3rd IEEE ISGT Europe, Berlin. Sidorov, D. (2015) Integral dynamical models: singularities, signals and control, vol. 87 of World Scientific Series on Nonlinear Science Series A. Singapore: World Scientific Tomin N., Negnevitsky M., Rehtanz Ch. (2014). Preventing Large-Scale Emergencies in Modern Power Systems: 450

AI Approach, Journal of Advanced Computational Intelligence and Intelligent Informatics, Vol. 18, No.5 Tomin N., Zhukov A., Sidorov A., Kurbatsky V., Panasetsky D., and Spiryaev V. (2015) Random forest based model for preventing large-scale emergencies in power systems, International Journal of Artificial Intelligence, Vol. 13, No. 1, pp. 211–228. Wang X., Shao W., Vittal V. (2005). Adaptive corrective control strategies for preventing power system blackouts in Proc. of 15th PSCC, Liege Wehenkel L. (1995). Machine Learning Approaches to Power System Security Assessment. Dissertation, University of Liege. Wehenkel, L., Van Cutsem, T., Pavella, M., Jacquemart, Y., Heilbronn, B., and Pruvot, P. (1994) Machine learning, neural networks and statistical pattern recognition for voltage security : a comparative study, Engineering Intelligent Systems for Electrical Engineering and Communications, Vol. 2 No 4, pp 233245. Wolpert, D. (1996) The Lack of A Priori Distinctions between Learning Algorithms, Neural Computation, pp. 1341-1390.