ontology matching system for future energy smart grids

Ontology matching system for future energy smart grids

R. Santodomingo a,n, S. Rohjans a, M. Uslar a, J.A. Rodríguez-Mondéjar b, M.A. Sanz-Bobi b

a OFFIS Institute for Information Technology, 2 Escherweg, Oldenburg 26121, Germanyb Institute for Research in Technology (IIT), Comillas Pontifical University, 23 Alberto Aguilera, Madrid 28015, Spain

a r t i c l e i n f o

Article history:Received 12 July 2013Received in revised form5 February 2014Accepted 7 February 2014Available online 4 March 2014

Keywords:Expert systemInteroperabilityKnowledge processingOntology matchingPower systemSmart grid

a b s t r a c t

Future power systems (commonly referred to as Smart Grids) will be managed by numerous intelligentelectronic devices. These devices will have to interoperate; that is, they will need to exchange data with eachother in order to co-operate over complex control tasks. Interoperability will only be achieved when SmartGrid devices share common semantics on the data they exchange. Standardization bodies have createdstandard data models defining these common semantics, but a unified standard data model has not beencreated for all Smart Grids. Consequently, in order to achieve interoperability in this domain, it is mandatoryto find semantic correspondences (alignments) between different standard data models. Creating equivalentontologies from the standard data models facilitates this task, because ontologies provide powerful reasoningservices that can be used for automating ontology aligning. The majority of ontology matchers proposed inthe state of the art, however, are only able to find simple equivalences of terms, while most alignments inSmart Grids are complex correspondences involving more than two terms. This paper presents an innovativeontology matching system that finds complex correspondences by processing expert knowledge fromexternal domain ontologies and by using novel matching methods. The tests carried out in this study werebased on the main interoperability issue within Smart Grids: interactions between CIM and SCL data models.In such tests, the proposed system outperformed one of the best ontology matchers according to theOntology Alignment Evaluation Initiative (OAEI).

& 2014 Elsevier Ltd. All rights reserved.

1. Introduction

The term Smart Grid is nowadays synonymously used forfuture power systems (Mc Namara et al., 2013; Rayudu, 2010).The former centralized infrastructure with its unidirectionalpower flow from large power plants via the consumers, turns intoa full-meshed topology including bidirectional power flows. Thisraises many challenges in terms of interoperability issues. Thus,numerous Smart Grid devices, possibly from different developers,need to exchange data with each other in order to co-operate overcomplex control tasks. The main standardization body in theelectricity sector, the International Electrotechnical Commission(IEC), has created some standard data models defining semanticsof the data that need to be exchanged within this domain.Unfortunately, a single standard data model has not been definedfor all Smart Grids. It is worth noting that many contributionsin the literature (Haslhofer and Klas, 2010; O'Leary, 1997) haveconcluded that, in practice, it is not always advisable (or evenpossible) to create a single standard data model that is valid for all

the applications within a domain. Therefore, with the aim ofachieving interoperability in Smart Grids, it is mandatory to aligndifferent data models.

Before aligning the models, these must be expressed in the samemodeling language. In this work, data models have been convertedinto OWL (Web Ontology Language) ontologies (Bechhofer et al.,2004). Originally, ontologies were created in Artificial Intelligence(AI) to produce knowledge base components for intelligent systems(Gómez-Pérez et al., 2004). More recently, ontologies have startedto be used in numerous engineering applications (Abanda et al.,2013; Blomqvist and Öhgren, 2008; Morbach et al., 2007; Wriggerset al., 2007). In this study, ontologies have been employed becausethey represent knowledge by defining logic theories, which enablemachines to infer implicit knowledge by means of general reasoningservices. These reasoning services are very useful for findingsemantic correspondences (alignments) between two data models.

The process of aligning ontologies is known as ontologymatching. Most ontology matchers are only able to find equiv-alences between ontology entities (Euzenat and Shvaiko, 2007).In Smart Grids, however, very few alignments can be expressed assimple equivalences (Santodomingo et al., 2012). For this reason,this paper presents a new ontology matching system that is able toobtain complex alignments between Smart Grid ontologies. This

Contents lists available at ScienceDirect

journal homepage: www.elsevier.com/locate/engappai

Engineering Applications of Artificial Intelligence

http://dx.doi.org/10.1016/j.engappai.2014.02.0050952-1976 & 2014 Elsevier Ltd. All rights reserved.

n Corresponding author. Tel.: þ49 441 9722 179; fax: þ49 441 9722 102.E-mail address: [email protected] (R. Santodomingo).

Engineering Applications of Artificial Intelligence 32 (2014) 242–257

www.sciencedirect.com/science/journal/09521976

www.elsevier.com/locate/engappai

http://dx.doi.org/10.1016/j.engappai.2014.02.005



http://crossmark.crossref.org/dialog/?doi=10.1016/j.engappai.2014.02.005&domain=pdf



mailto:[email protected]


system processes deep domain knowledge from external ontolo-gies (domain ontologies) and uses innovative matching methods.

The remainder of this paper is organized as follows. Section 2provides a brief overview of the two main standard data models thatpromote interoperability in Smart Grids: CIM and SCL. Section 3 setsout the fundamentals aspects of Ontology Matching. Section 4presents the proposed ontology matching system, which comprisesa schema-based subsystem (Section 5) and an instance-basedsubsystem (Section 6). Section 7 is devoted to the tests performedin this study to evaluate the proposed system. Finally, Section 8concludes the paper and provides an outlook on future work.

2. Interoperability in smart grids

This work addresses the main interoperability issue in SmartGrids: interactions between CIM-based and SCL-based systems.This section presents CIM and SCL data models and discusses theirinteractions.

2.1. Smart grid data models: CIM and SCL

The Common Information Model (CIM) is defined in the IEC 61970/61968/62325 standard series. It standardizes the semantics toachieve interoperability in a broad range of energy managementfunctionalities, such as: network operations, asset management andelectricity markets (Uslar et al., 2012). Hundreds of classes organizedin packages are included in this data model. Among all the CIMpackages, Wires and Topology packages contain several classes torepresent electric power systems. For instance, cim:Substation,cim:VoltageLevel, and cim:Breaker are the CIM classes torepresent substations, voltage levels, and circuit breakers, respec-tively. The CIM is currently maintained in UML. In this study, theequivalent CIM OWL ontology has been created by using theUml2Owl conversion implemented in the open-source CIMTool.1

The Substation Configuration Language (SCL) is defined in theIEC 61850 standard series. It includes the concepts required forconfiguring the automation systems that locally control electricnetworks. The SCL defines terms to represent automation systemsand electric facilities. For instance, scl:tSubstation and scl:tVolta-geLevel are the SCL classes for representing substations andvoltage levels, respectively. Meanwhile, circuit breakers are repre-sented in SCL as scl:tConductingEquipment instances thattake the value “CBR” in the attribute scl:type. The SCL isrepresented in the XML Schema Definition (XSD) language. Theequivalent SCL OWL ontology has been created in this work withthe Xsd2Owl conversion presented in (García and Gil, 2007).

2.2. Interactions between CIM and SCL

As detailed in (Falk and Saxton, 2010), during the planning andconfiguration of electric networks, engineers of the CIM-basedmanagement system must exchange files with engineers of theSCL-based automation system. The problem is that CIM and SCLwere developed by different working groups with differentrequirements and goals. This resulted in the existence of mis-matches hindering the interactions between CIM-based and SCL-based systems. In that way, simple equivalences of terms are notsufficient themselves for carrying out these interactions. Forinstance, CIM and SCL representations of circuit breakers mustbe aligned by a complex correspondence involving the class

cim:Breaker, the class scl:tConductingEquipment, the attri-bute scl:type and the attribute value “CBR”.

3. Ontology matching

This section defines the concepts that will be used throughoutthe paper to describe the proposed ontology matching system. Forthat purpose, the terminology defined in (Euzenat and Shvaiko,2007) is employed.

3.1. Ontology matching process

Ontology matchers are aimed at finding correspondencesbetween entities (classes, properties, instances) of two ontologieso1 and o2. In this study, two types of property are considered:object properties (or relationships), which connect two classes orinstances, and data properties (or attributes), which connect aclass or instance to a data type (e.g., string, integer, etc.) or value.

Typically, ontology matching processes comprise two overallsteps: similarity computation and alignment extraction (Fig. 1).In the first step, entities of o1 and o2 are compared. For eachentity e1 of o1 and e2 of o2, a similarity measure s(e1,e2) iscalculated. This measure is a function from a pair of entities to areal number expressing the similarity between them (Euzenat andShvaiko, 2007). In order to work with similarity measures withinthe range [0–1], usually, similarities are normalized with themaximum s(e1,e2) that was calculated. The results obtained in asimilarity computation are included in a similarity matrix Mcontaining all the similarity measures between entities of o1 andentities of o2. Several similarity computations are proposed in thestate of the art. These can be categorized in

� Element-based methods, which analyze the entities of the ontol-ogies in isolation (i.e., ignoring their relationships with otherentities). In this work, different element-based methods includedin the state of the art (such as the string-based comparisonproposed in (Winkler, 1999)) were used (see Section 5).

� Structure-based methods, which take the structure of the ontologies(i.e., the relationships between entities) into account. Graph-basedmatching methods were identified as the most appropriatestructure-based methods for the system proposed in this work(see Section 6.3.3). This is because graph-based methods aremature methods that obtain good results in the ontology align-ment evaluations (Cruz et al., 2011). These methods represent

o1 o2

AlignmentExtraction

SimilarityComputation

A

M

Fig. 1. Ontology matching process.

1 CIMTool (from Langdale Consultants) is available at: http://wiki.cimtool.org/Download.html (accessed on January 2014).

R. Santodomingo et al. / Engineering Applications of Artificial Intelligence 32 (2014) 242–257 243

http://wiki.cimtool.org/Download.html

http://wiki.cimtool.org/Download.html

ontologies as graphs and compare their nodes (classes or instances)attending to local characteristics (such as attributes or names).The local similarities s� l(e1,e2) obtained from this comparison arethen propagated by considering the structure of the graphs inorder to obtain the global similarities s�g(e1,e2). The main graph-based matching methods proposed in the literature are: theSimilarity Flooding (SF) defined in (Melnik et al., 2002); theSimilarity Equation Fixed Point (SEFP) defined in (Euzenat andValtchev, 2004); the Descendant's Similarity Inheritance (DSI) andthe Sibling's Similarity Contribution (SSC) defined in (Cruz andSunna, 2008); and the Semantic-Neighbourhood Matching (SNM)defined in (Rodríguez and Egenhofer, 2003).

Existing mismatches between ontologies make for difficult useof a single similarity computation. Hence, the final similaritymatrix M is usually calculated from the similarity matrices Miobtained with different similarity computations (Shvaiko andEuzenat, 2013). Two approaches have been used in this work tocombine similarity computations: the ontology-driven combina-tion, which calculates the weights of the matrices Mi by analyzingthe ontologies for instance, the combination proposed in (Juanziet al., 2009), and the quality-driven combination defined in (Cruzet al., 2009), which calculates the weights by analyzing the qualityof the similarity matrices Mi.

In the second step (alignment extraction), the alignments betweenontologies are obtained from the final similarity matrix M. Theliterature includes two basic techniques to carry out the alignmentextraction: threshold-based extraction and mapping-based extraction.The threshold-based extraction selects the pairs of entities ⟨e1,e2⟩whose normalized similarity is greater than a configurable threshold.For each pair of entities selected, an equivalence e1� e2 is created. Themapping-based extraction uses mapping algorithms that analyze thesimilarity matrix M to select the pairs of entities ⟨e1,e2⟩ that areequivalent. Therefore, both techniques are only able to obtain equiv-alences between ontology entities. This partially explains why mostontology matchers are not able to obtain complex correspondences.The main mapping algorithms proposed in the literature are: thePerfect Monogamy (PM) and the Stable Marriage (SM) defined in (Marieand Gal, 2007); and the Maximum Weight Bipartite Graph (MWBG)defined in (Munkres, 1957).

3.2. Matching with background knowledge

Apart from the classification of similarity computations explainedin Section 3.1 of this contribution, ontology matching systems can becategorized also by considering the type of information they use foraligning ontologies. Hence, ontology matching systems can employ:syntax of the elements included in the ontologies to be aligned, formalsemantics of the ontologies to be aligned, or background knowledgefrom external resources (Euzenat and Shvaiko, 2007).

In their recent review of the state of the art, Shvaiko and Euzenatexplain that “one source of difficulty for matching is that ontologiesare designed in a particular context, with some background knowl-edge, which often do not become part of the final ontology specifica-tion”. Thus, the authors point out that “matching with backgroundknowledge” is one of the most important challenges for ontologymatching (Shvaiko and Euzenat, 2013). Several studies have analyzedhow to infer background knowledge from external ontologies toimprove ontology alignment. Aleksovski explored the influence ofdifferent types of background ontologies (Aleksovski, 2008). Mascardiet al. proposed algorithms to exploit upper ontologies (that is,ontologies defining general concepts that are the same across alldomains) as background knowledge for ontology matching (Mascardiet al., 2010). Finally, Sabou et al. defined a paradigm to automaticallyfind background ontologies in the Semantic Web (Sabou et al., 2008).In all these contributions, the process of matching with background

knowledge comprises two steps: anchoring and deriving relations(Fig. 2).

� Anchoring refers to the mapping of entities from the source andtarget ontologies to entities of the background ontologies. Thismapping follows the generic ontology matching processdescribed in Section 3.1.

� Deriving relations is the process of discovering relationsbetween source and target entities by looking for relationsbetween their anchored concepts in the background knowl-edge (Aleksovski, 2008). In the solutions proposed in the stateof the art, derived relations are simple correspondences(equivalence, subsumption, or disjointness) between one entityof the source ontology and one entity of the target ontology.

The ontology matching system proposed in our work uses back-ground knowledge to align ontologies in the Smart Grids domain.However, unlike existing solutions, our system performs the anchoringby mapping instances, previously translated from the source to thetarget ontology (d11), to concepts of domain ontologies. As will bedetailed in Section 6, this process enables to automatically findcomplex correspondences relating more than two entities (Fig. 3).

3.3. Ontology alignment evaluation

Starting from 2004, the Ontology Alignment Evaluation Initiative(OAEI)2 campaign is run every year with the purpose of establishinga consensus for evaluating ontology matching methods. This initia-tive uses two evaluation measurements: compliance measurementsand performance measurements. Compliance measurements

o1 o2

Background knowledge ontologies

Anchoring AnchoringDerivingrelations

Inferred simple correspondences

Fig. 2. Typical matching with background knowledge, from (Aleksovski, 2008).

o1 o2

Background knowledge ontologies

(domain ontologies)

Anchoring Derivingrelations

Inferred complex correspondences

d1º

Instance of Instance of

Fig. 3. Matching with background knowledge in our system.

2 Ontology Alignment Evaluation Initiative (OAEI): http://oaei.ontologymatching.org/.

R. Santodomingo et al. / Engineering Applications of Artificial Intelligence 32 (2014) 242–257244

http://oaei.ontologymatching.org/

compute the quality of the output alignments provided by theontology matchers in comparison with the reference alignments,which comprise all the valid correspondences between the ontolo-gies that have to be aligned. The most common compliancemeasurements in ontology matching are: Recall, Precision and F-measure (Euzenat and Shvaiko, 2007). Recall measures the ratio ofcorrectly found correspondences (true positives) to the total numberof reference alignments; Precision measures the ratio of truepositives to the total number of returned correspondences; and F-measure is the harmonic mean of recall and precision. Meanwhile,performance measurements refer to the runtimes (in seconds) of thetests performed with different ontology matchers.

4. Proposed ontology matching system

The ontology matching system proposed in this paper consistsof a schema-based subsystem and an instance-based subsystem(Fig. 4).

The schema-based subsystem only uses schema-level inputinformation. It imports the standard ontologies (o1 and o2) andcreates the initial alignments (A1) between them by combiningdifferent element-based methods. This sub-system also takesadvantage of an electrical and electronic terminology databasecalled electropedia.3

For its part, the instance-based subsystem takes advantage ofinstance files. It involves an iterative process between the ESODATdata translator presented in (Santodomingo et al., 2012) and an expertsystem developed in this study, the Domain Expert. By processing theinitial alignments (A1), ESODAT translates an instance file (d1) from thesource ontology (o1) into the target ontology (o2). Given that thealignments do not include all the existing correspondences betweenthe ontologies, there will be data losses in the output of the translation(d11). The Domain Expert automatically detects data losses in d11 andproposes new alignments (A0) for improving the translation in the nextiteration. In order to do this, the Domain Expert employs its knowledgebase previously created from domain ontologies that formally expressdeep background knowledge about representative substation archi-tectures. The iterative process continues until the Domain Expertdetects that the translation has not improved from the previousiteration.

It is worth noting that Fig. 4 details the process for finding thecorrespondences required to translate files from o1 to o2. Whenthe correspondences in the opposite direction are required, it isnecessary to repeat the process by importing an instance file (d2)coming from the ontology o2.

Fig. 5 uses the classification proposed in (Euzenat and Shvaiko,2007) to summarize the ontology matching methods employed inour schema-based and instance-based subsystems. These methodswill be explained in Sections 5 and 6.

5. Schema-based subsystem

The schema-based subsystem is implemented in the CIMMap-pingBench, which is a stand-alone tool defined in (Uslar et al.,2008). It combines the following matching methods: the linguisticmethod, the language-based method, the string-based methodand the constraint-based method.

In the linguistic method, an equivalent ontology was createdfrom the electropedia terminology database in order to identifywords with similar meanings and in different languages. Thisontology is used to complete the strings of the entities by addingthe definitions given in the electropedia. The extended strings of

the entities are then prepared with language-based methods. First,a word stemming algorithm (Porter, 1980; Tartarus, 1980) reducesall the words within the names and descriptions to their root form,and next, a file including stop words (Cornell, 1999) is used to filterout redundant words.

Once the strings of the entities have been prepared, string-based and constraint-based methods are used to compare entitiesof the ontologies o1 and o2. The string-based method comparesthe strings of the entities using the Jaro-Winkler algorithm(Winkler, 1999). This comparison results in the similarity matrixMstring, whose elements s-string (e1,e2) give a unitary measure ofthe similarity between the entities e1 and e2. For its part, theconstraint-based method compares the attributes of the entities;so it is calculated only for the similarities between classes. Thismethod assumes that two classes can be equivalent (even if theirnames and descriptions are different) if more than 80% of theirattributes are similar. Two attributes a1 and a2 are consideredsimilar if their corresponding s-string (a1,a2) value is greater than0.9. The calculated similarity in Mconstraint (s-constraint) issteadily increasing starting at 0 for 0% similar attributes andending at 1 for 100% similar attributes.

Then, the CIMMappingBench combines the matrices Mstringand Mconstraint to obtain the final similarity matrix M. When thes-constraint is greater than 0.8 (that is, the classes have similarsets of attributes) the element s of the similarity matrix M iscalculated as the arithmetic mean of s-string and s-constraint.Otherwise, s is simply the s-string value.

Finally, the schema-based subsystem uses a threshold-basedextraction method to obtain the initial alignments (equivalences) fromthe final similarity matrixM. These alignments must be very precise in

string-based constraint-based

language-based

combination

o1 o2

Mstring Mconstraint

Aº

electropedia

M

Schema-based subsystem

data translatord1

domainontologies

domain expertd1ºIterationº

A

A’

data translator domain expertd1’Iteration’

data translator domain expert

Iterationn

An

Instance-based subsystem

d1

d1d1n

linguistic

threshold-basedextraction

Fig. 4. Ontology matching system for Smart Grids.

3 IEC 60050: Electropedia, IEC TC1, available at: http://www.electropedia.org/.


http://www.electropedia.org/

order to facilitate convergence of the iterative process in the instance-based subsystem. Hence, in this study, the threshold was set in theCIMMappingBench at 0.85, which is a very restricted value.

6. Instance-based subsystem

The instance-based subsystem comprises two processes: thedata translation and the ontology matching process. The datatranslation is carried out by the ESODAT data translator presentedin (Santodomingo et al., 2012). This section focuses on theontology matching process, which is performed by the expertsystem developed in this work, the Domain Expert.

Before describing this ontology matching process, it should benoted that the instance-based subsystem filters the initial align-ments imported from the schema-based subsystem in the firstiteration. This filtering process removes equivalences betweenclasses and properties (e.g., between the class cim:VoltageLeveland the property scl:VoltageLevel). Moreover, when a class isequivalent to two classes of the other ontology, it removes theequivalence with lower similarity.

6.1. Domain ontologies

The Domain Expert creates its knowledge base by importingdomain ontologies. This knowledge base enables the DomainExpert to detect data losses that occur during data translations inorder to find new alignments. What follows explains the reasonsfor creating domain ontologies in Smart Grids and gives anoverview of the domain ontologies developed in this work.

CIM and SCL define classes that can be used to representelectric facilities. However, these standard models do not provideprecise definitions for describing a particular type of electricfacilities. For instance, CIM includes the class cim:Substation

to represent electrical substations, but it does not define thecharacteristics of a typical radial substation, e.g.,: “it must containtwo voltage levels”, “one of the voltage levels must contain anelectrical busbar”, etc. This means that it is possible to create validCIM and SCL files representing electric facilities that can rarelyexist in practice, such as a substation that only consists of onecircuit breaker. Consequently, the Domain Expert cannot evaluatedata translations by importing the standard CIM and SCL ontolo-gies. Therefore, it is mandatory to create domain ontologies thatformally describe how to represent typical electric facilities in CIMand SCL. Given that there are no many types of electric facilities inpractise, in a realistic scenario it should be possible to create adatabase containing all the domain ontologies that are required forrepresenting complete power systems.

In this work, the domain ontologies have been created in OWLand SWRL (Semantic Web Rule Language) (Horrocks et al., 2004)by using the open-source ontology editor Protégé 3.4.6.4 Thesedomain ontologies describe how to represent in CIM and SCL fiverepresentative substation architectures that were considered inthe case studies (Section 7.1). The overall structure of such domainontologies is presented next with a simple example. Fig. 6 showsan extract of the domain ontology that describes how to representin CIM the H topology (or Type_3), which is one of the substation

elem

ent-b

ased

mat

chin

g

Syn

tax

of

elem

ents

Ext

erna

lR

esou

rces

String-basedmethods Jaro-Winkler

stru

ctur

e -ba

sed

mat

chin

g

Ontology Matching Methods Schema-based subsystem Instance-based subsystem

Language-basedmethods

Constraint-basedmethods

Linguistic methods

Using existingalignments

Using externalOntologies

Data analysis and statistics

Graph-basedmethods

Taxonomy-basedmethods

Syn

tax

of

elem

ents

repository of structures

Form

al

Sem

ant.

Ext

. R

es.

model-basedmethods

StemmingStop words

Sets of data properties

Electropedia

Importing initialalignments

MSC

Domainontologies

Sets of data and object properties

Fig. 5. Ontology matching methods used in the proposed ontology matching system.

4 Protegé 3.4.6 is available at: http://protege.cim3.net/download/old-releases/Protege%203.x/3.4.6/ (accessed on July 2013).


http://protege.cim3.net/download/old-releases/Protege%203.x/3.4.6/

http://protege.cim3.net/download/old-releases/Protege%203.x/3.4.6/

architectures given in the case studies. As can be observed, thisextract defines three classes: CIMHSubstation, CIMD1 and CIME1.

CIMHSubstation derives from cim:Substation. With the aimof describing the particularities of H topologies, CIMHSubstationadds two constraints establishing that this class must have exactlyone CIMHSubstation-CIMD1 property and one CIMHSubstation-CIME1property. Such object properties derive from cim:VoltageLevel.

Substation, which is the CIM property used to represent therelationships between substations and voltage levels. Therefore, theconstraints mentioned above establish that H topologies must havetwo voltage levels.

CIMD1 and CIME1 derive from cim:VoltageLevel. Theseclasses define the particular characteristics of the two voltagelevels that must be contained in H topologies. Thus, the definitionsof CIMD1 and CIME1 (which, for clarity purposes, are not includedin Fig. 6) provide detailed descriptions of the bays and conductingequipment (e.g., circuit breakers, and disconnectors) that shouldbe included in these voltage levels. In this way, the domainontologies created in this work define all the classes and proper-ties that are required to describe in CIM and in SCL five repre-sentative substation architectures.

6.2. Ontology matching process in the domain expert

Ontology matching processes typically obtain the alignmentsby importing the standard ontologies that have to be aligned(Fig. 1). Nonetheless, in the Domain Expert, the ontology matchingprocess is carried out by importing an instance file d11 (output ofthe data translation) and two domain ontologies: the sourcedomain ontology and the target domain ontology (Fig. 7).

The Domain Expert first compares d11 with the source andtarget domain ontologies by using innovative similarity computa-tions (Section 6.3). As a result of these comparisons, two similaritymatrices Ms and Mt are obtained, which establish the similaritiesbetween the instances in d11 and the classes in the source and thetarget domain ontologies, respectively. From Ms and Mt, alignmentextraction methods developed in this work (Section 6.4) obtaincomplex alignments between the standard ontologies.

6.3. Similarity computation

As discussed in Section 3.1, the similarity computations devel-oped in this study are graph-based matching methods. Section6.3.1 describes how the Domain Expert creates graphs from d11 andfrom the domain ontologies. Then, Section 6.3.2 details how these

graphs are compared by considering local characteristics of thenodes. Finally, Section 6.3.3 presents the innovative graph-basedmatching algorithms that propagate the local similarities in orderto obtain global similarities.

6.3.1. Knowledge acquisitionThe knowledge base created in the Domain Expert comprises three

graphs: I, Ds and Dt. The first graph is created from d11. Its nodes irepresent the instances described in the output of the translation,whereas the arcs that connect the nodes represent object propertiesbetween the instances. Graphs Ds and Dt are created from the sourcedomain ontology and the target domain ontology, respectively. Theirnodes d represent domain ontology classes, whereas the arcs repre-sent object properties between the domain ontology classes.

The following example is used to describe the graphs mentionedabove. In the example, the objective is to obtain the correspon-dences that are required to carry out translations from CIM into SCL.Let us suppose that the instance-based subsystem uses an existingCIM file describing an electric facility that consists of a bay (E1Q1)containing a circuit breaker (QA1). Moreover, let us suppose that theinitial alignments found by the schema-based subsystem (A1)include the correspondences between cim:IdentifiedObject.

name and scl:name and also between cim:Bay and scl:tBay.In that way, the output of the translation d11 contains some SCLterms that are inferred in the data translator (ESODAT) by processingthe initial alignments. For example, E1Q1 will be represented in d11as an instance of cim:Bay, but also as an instance of scl:tBay.Finally, let us suppose that the source domain ontology and thetarget domain ontology imported in the Domain Expert describe howto represent this type of electric facility in CIM and SCL, respectively.Fig. 8 shows the graphs I and Dt created in this example.

In graph I, nodes include following two elements:

� Classes represent the classes of the instances in the output ofthe translation. For example, the node E1Q1 belongs to twoclasses: cim:Bay and scl:tBay.

� Attributes refer to the data properties that appear in thedefinition of the instance in the output of the translation. Forexample, the node E1Q1 has the attributes scl:name and cim:

Identified Object.name.

Meanwhile, in graph Dt (the same as in graph Ds) nodescontain the following elements:

� Standard Class represents the class of the standard ontologyfromwhich the domain ontology class derives. For instance, thedomain ontology class SCLBay1 represents a particular type ofbay, so it derives from the standard ontology class scl:tBay.

CIMHSubstation-CIME1

CIMHSubstation-CIMD1

DomainOntology

CIM Ontology

has subclasshas subclasshas subclass

Fig. 6. Domain ontology describing the H topology (Type_3) in CIM.

source domainontology


d1º

target domainontology


Alignment Extraction

Ms Mt

A

Fig. 7. Ontology matching process in the Domain Expert.


� Fixed Value Attributes refer to the constraints in the definitionsof the domain ontology classes that set a specific value for anattribute. For example, in SCL, circuit breakers are represented withscl:tConductingEquipment instances that take the value“CBR” in the attribute scl:type. Thus, the domain ontology classSCLBreaker1 includes the fixed value attribute scl:type¼“CBR”.

� Reference Attributes are the attributes used to establish relation-ships by reference; that is, relationships between two classesthat are related when they take the same value in theirreference attributes. No relationships by reference are includedin the example. However, in SCL, connections between term-inals and connectivity nodes are represented with relationshipsby reference based on the reference attributes scl:connec-

tivityNode and scl:pathName.� Attributes refer to all the other attributes that appear in the

definition of the domain ontology class.

Once the graphs have been created, the Domain Expert com-pares the nodes of graph I with the nodes of graphs Ds and Dt.These comparisons are performed with the ontology matchingmethods presented in the following sub-sections. It should benoted that the explanation of these methods is based on thecomparison between graph I and a generic graph D. However, asshown in Fig. 7, this process is carried out for both Ds and Dt.

6.3.2. Local similaritiesIn order to compare graph I with graph D, the Domain Expert

first obtains the local similarities s� l(i,d). These local similaritiesgive a unitary measurement of the similarity between the nodes iof I and the nodes d of D focusing on their local characteristics.In this context these characteristics are: name of the correspond-ing standard ontology class; label of the arcs connected to thenodes; and name and value of the attributes included in the nodes.Hence, the local similarity matrix Ml containing all the similaritiess� l(i,d) is obtained in this work as the weighted sum aggregation ofMc, Mr and Ma (1), with:

� Mc being the matrix containing the similarities s�c(i,d) betweenthe names of the standard ontology classes related to i and d,

� Mr being the matrix containing the similarities s�r(i,d) betweenthe labels of the arcs (i.e., object properties) connected to i in Iand to d in D,

� and Ma being the matrix containing the similarities s�a(i,d)between the attributes of i and d.

Ml¼wc �Mcþwr �Mrþwa �Ma ð1Þ

In Mc, s�c(i,d) is 1 if i belongs to the standard ontology classfrom which d derives, and is 0, on the contrary. For instance, inFig. 8, the value of s�c(E1Q1,SCLBay1) is 1, whereas the value ofs�c(E1Q1,SCLBreaker1) is 0.

Meanwhile, in Mr, elements s�r(i,d) are calculated with theexpression given in (2), with: Ri being the set of all the arcsconnected to i in I, Rd being the set of all the arcs connected to d inD, and | | being the number of elements contained in a set. Forinstance, in Fig. 8, s�r(E1Q1,SCLBay1) is calculated as follows. Theterm |Ri\Rd| is the intersection between the set of arcs connectedto E1Q1 and the set of arcs connected to SCLBay1. These nodeshave no arcs in common. Therefore, in this case, |Ri\Rd|¼s�r(E1Q1,SCLBay1)¼0.

s�rði; dÞ ¼ jRi \ Rdj=jRi [ Rdj ð2Þ

finally, elements s�a(i,d) of Ma are calculated with the expressiongiven in (3), with: Ai being the set of attributes included in i, andAd being the set of attributes included in d. Following the examplepresented above, s�r(E1Q1,SCLBay1)¼1/1¼1, because the onlySCL attribute in E1Q1 and SCLBay1 is scl:name.

s�aði; dÞ ¼ jAi \ Adj=jAi [ Adj ð3Þ

As stated in Section 3.1, two main approaches proposed in thestate of the art have been considered in this work for combiningsimilarity computations: ontology-driven approach and quality-driven approach. What follows explains why the ontology-drivenapproach was used in this study to calculate the weights wc, wrand wa in Eq. (1).

E1Q1Classes:

cim:Bayscl:tBay

Attributes:cim:IdentifiedObject.name = “E1Q1”scl:name = “E1Q1”

SCLBay1Standard Class:

scl:tBay

Attributes:scl:name

SCLBreaker1Standard Class:

scl:tConductingEquipment

Fixed Value Attributes:scl:type = “CBR”

Attributes:scl:name

QA1Classes:

cim:Breaker

Attributes:cim:IdentifiedObject.name = “QA1”scl:name = “QA1”

cim:Equipment.EquipmentContainer

scl:Conducting-Equipment

Graph I Graph Dt

Fig. 8. Graphs I and Dt.


In order to explain this choice, two aspects must be taken intoaccount: (a) in our case, each entity i of I can only be mapped to, atthe most, one entity d of D; and (b) for each node i there is usuallymore than one node d with the same standard ontology class in D.The second aspect implies that, typically, in the matrix Mc therewill be instances i taking the similarity value 1 for differentdomain ontology classes d. Considering the first aspect mentionedbefore, and using the quality parameters defined in (Cruz et al.,2009), it can be concluded that matrix Mc will have usually a lowquality. Therefore, in most cases, the weights wc calculated by thequality-driven approach will be very low, or even 0. Despite of itslow quality, Mc provides very useful information for carrying outthe subsequent analyses (Section 6.3.3 and Section 6.4), because itfinds which instances and domain ontology classes have the samestandard ontology class, i.e., the same type. Hence, in order to takeadvantage of the information provided by Mc, parameters otherthan the quality of the similarity matrices should be used forcombining similarity computations in this case.

These parameters can be given by analyzing the graphs to bemapped (ontology-driven approach). In particular, the ontology-driven method employed in this study to calculate the weights wc,wr and wa, was the one proposed in (Juanzi et al., 2009). Theoverall intuition behind this ontology-driven combination can beexpressed as follows: “the more information about the localcharacteristics considered in a similarity matrix, the greater theweight assigned to that matrix”. As it will be explained next, thismethod is very useful in our context. Let us suppose that the initialalignments A1 have more correspondences between classes thanbetween properties and attributes. In this scenario, during the firstiteration graph I will have more relevant information about classesthan about relationships and attributes, and, therefore, Mc will bemore useful than Mr and Ma for finding the mappings. This fitsperfectly with the ontology-driven method proposed in (Juanziet al., 2009), as it will assign a greater value to wc than to wr andwa. When in the first iteration the system finds many correspon-dences between relationships, in the second iteration the weightwr calculated by the ontology-driven method will grow, becausethe matrix Mr will provide now more meaningful information.In that way, the ontology-driven method sets the weights to bettercombine the similarity matrices at each step of the iterativeprocess.

6.3.3. Global similaritiesWith the aim of obtaining the global similarities it is necessary

to propagate the local similarities by considering the overallstructures of the graphs. As discussed in Section 3.1, existinggraph-based matching methods proposed in the literature can beused for this purpose. This section presents two innovative graph-based matching methods developed in this study: Descendants'Similarity Contribution (DSC) and the Mixed Similarity Contribution(MSC).

The Descendants' Similarity Contribution (DSC) is a new algo-rithm based on the Descendant's Similarity Inheritance algorithm(DSI) proposed in (Cruz and Sunna, 2008). Whereas in DSI,containers propagate their similarities to their children in thegraph, in the DSC, children propagate their similarities to theircontainers. What follows explains how the DSC calculates theglobal similarity between the instance i and the domain ontologyclass d. Let path_len_leave(i) be the minimum number of arcsbetween i and the farthest instance contained in i, and path_len_leave(d) be the minimum number of arcs between d and thefarthest domain ontology class contained in d. From these values,the DSC calculates the factors n, which is the maximum value ofpath_len_leave(i) and path_len_leave(d), and n2, which is theminimum value of path_len_leave(i) and path_len_leave(d). Once

the factors n and n2 are obtained, the global similarity s�gDSC(i,d)is calculated with the expression given in (4).

s�gDSCði; dÞ ¼MCP �s� lði; dÞþ2 � ð1�MCPÞn � ðnþ1Þ � ∑

n2�1

m ¼ 0ðn2�mÞ �MSimðSim; SdmÞ

ð4Þ

In (4):

� MCP is the main contribution percentage, which is the fractionof the local similarity s� l(i,d) that will be used in determiningthe global similarity s�gDSC(i,d). As in (Cruz and Sunna, 2008),in this study the MCP was set at 0.75.

� Sim and Sdm are the set of nodes contained in i (respectively, d)that have m arcs between them and i (respectively, d).

� MSim is the match-based similarity between two sets ofentities; in this case, between Sim and Sdm. As explained inchapter 4 of (Euzenat and Shvaiko, 2007), MSim(x,y) is obtainedwith the expression given in (5); with Pairings(x,y) being theset of mappings between x and y.

MSimðx; yÞ ¼maxpAPairingsðx;yÞð∑on;n0 4 Aps� lðn; n0ÞÞmaxðjxj; jyjÞ ð5Þ

Fig. 9 shows how the DSC obtains the global similarity betweenthe instance E1Q1 and the domain ontology class SCLBay1. In theexample, MSim�1 and MSim�2 measure the similarity betweenthe sets {L1, QA1} and {SCLConnectivityNode1, SCLBreaker1}, andbetween the sets {T1} and {SCLTerminal1}, respectively. In thiscase, path_len_leave(i) and path_len_leave(d) are both equal to 2,because the shortest path that connects T1 (respectively, SCLTerm-inal1) to E1Q1(respectively, SCLBay1) has 2 arcs. Therefore, para-meters n and n2 are both equal to 2, and, using the expressiongiven in (4), the weights assigned to MSim�1 and MSim�2 resultin 0.167 and 0.083, respectively.

The Mixed Similarity Contribution (MSC) combines the simila-rities calculated with the DSI, the SSC (which is also proposed in(Cruz and Sunna, 2008)), and the DSC. Thus, the global similaritys�gMSC(i,d) is obtained with the equation given in (6).

s�gMSC ¼wDSI �s�gDSIþwSSC �s�gSSCþwDSC �s�gDSC ð6Þ

The three methods DSI, SSC and DSC use the same ontologyentities to obtain global similarities, so in this case it is notpossible to calculate the weights wDSI, wSSC and wDSC by analyzingthe ontologies (graphs). Consequently, the MSC developed in thiswork employs the quality-driven combination proposed in (Cruzet al., 2009). This means that the weights wDSI, wSSC and wDSC areautomatically obtained by analyzing the quality of the similaritymatrices MDSI, MSSC and MDSC.

MSim-2

MSim-1

σ-l(E1Q1,SCLBay1)

σ-gDSC(E1Q1,SCLBay1) = 0.75*σ-l(E1Q1,SCLBay1) + 0.167*MSim-1 + 0.083*MSim-2

E1Q1

QA1L1

T1

SCLBay1

SCLBreaker1SCLConnectivityNode1

SCLTerminal1

Fig. 9. Descendant's Similarity Contribution (DSC).


6.4. Alignment extraction

Once the global similarity matrices between I and Ds (i.e., Ms) andbetween I and Dt (i.e., Mt) have been obtained, it is possible to extractnew alignments for the next iteration. The alignment extractionmethod developed in this study, unlike most existing methods inthe literature, is able to find complex semantic correspondences. Thismethod comprises two steps. In the first step, from the globalsimilarities Ms and Mt, an innovative mapping algorithm maps theinstances of I to their most similar domain ontology classes in Ds andDt (Section 6.4.1). In the second step, from these mappings, thealignment extraction method detects losses that occurred in thetranslation and proposes new alignments with the aim of avoidingthese data losses in the next iteration (Section 6.4.2).

6.4.1. Mapping algorithmIn order to develop a mapping algorithm for the Domain Expert,

it is important to consider the following aspects:

� Every instance must be mapped to, at the most, one domainontology class.

� Every domain ontology class must be mapped to, at the most,one instance.

� Due to the existence of coverage mismatches (Euzenat andShvaiko, 2007), the mappings may not be bijective.

� The alignments created from the mappings obtained by themapping algorithm are used in the next iteration for translatingan instance file.

Therefore, the mapping algorithm must meet two requirements:

� It must establish [0,1] – [0,1] mappings (using the notation of(Melnik et al., 2002)).

� It must be precise; i.e., it should avoid the creation of falsemappings, because this could make future translations worse.

All the mapping algorithms mentioned in Section 3.1 (i.e., PM,SM and MWBG) meet the first requirement. As regards the secondrequirement, the precision of these algorithms can be improved byusing a threshold; that is, the mapping algorithm first obtains themappings, and then it only selects those mappings with a globalsimilarity greater than the threshold.

With the aim of enhancing the performance of the alignmentextraction, a new mapping algorithm has been developed in thiswork: the Hierarchical Mapping (HM). The HM is based on theperformance tuning detailed in (Cruz and Sunna, 2008). Fig. 10explains the HM algorithm by using a simple example. The HMstarts by comparing the global similarities between the immediatechildren of the root nodes of I and D. The best possible match foreach of the children is determined by the MWBG algorithm. Then theHM compares only the sub-trees of the mapped nodes. This meansthat if i has been mapped to d in the previous step, then only thechildren of i are compared with the children of d. This process isrepeated until the HM reaches the leave nodes.

6.4.2. Creation of complex correspondencesFrom the mappings between instances and domain ontology

classes, the Domain Expert is able to detect data losses that occurduring the data translation. In particular, four types of losses havebeen considered in this study (Table 1). As will be discussed at theend of this sub-section, Table 1 does not provide a comprehensivelist containing all the possible data losses that can occur duringdata translations. On the contrary, these are the losses that can beresolved by the Domain Expert by proposing new alignments.In future work, the Domain Expert should be able to detect moredata losses in order to find more alignments.

For each data loss, the Domain Expert proposes a complexalignment that resolves such a loss in the next iteration. In whatfollows, the four types of data losses and the correspondingcomplex alignments are described by using simple examples.

Data loss 1 – missing target class occurs when:

� The instance i is mapped in Ds to ds and in Dt to dt.� The instance i does not belong to the standard class of dt.

In this case, the alignment extraction method creates a SWRLalignment that establishes a transformation rule:

� From the standard class and fixed value attributes of ds.� To the standard class and fixed value attributes of dt.

For example, let us suppose that the instance QA1 has beenmapped to CIMBreaker1 in Ds and to SCLBreaker1 in Dt. As shownin Fig. 11, during the translation it was not inferred that theinstance QA1 belongs to scl:tConductingEquipment, which isthe standard class of SCLBreaker1. Therefore, the alignment extrac-tion method creates the SWRL alignment given in (7).

cim : Breakerð?xÞ-scl : tConductingEquipmentð?xÞ ∧

scl : typeð?x; }CBR}Þ ð7Þ

Data loss 2 – missing target relationship (equivalent to sourcerelationship) occurs when:

� The instance i is mapped in Ds to ds and in Dt to dt.� The instance in is mapped in Ds to dsn and in Dt to dtn.� In Ds there is a relationship rs(ds,dsn).� In Dt there is an equivalent relationship rt(dt,dtn).� In the output of the translation, i is not related to in by the

relationship rt(i,in).

In this case, the alignment extraction creates a SWRL alignmentthat establishes a transformation rule:

� From the standard class of ds and the relationship rs.� To the relationship rt.

For example, in Fig. 12, the instance E1 is mapped to CIMVolta-geLevel1 in Ds and to SCLVoltageLevel1 in Dt. Meanwhile, theinstance E1V is mapped to CIMVoltage1 in Ds and to SCLVoltage1in Dt. Moreover, E1 is connected to E1V with cim:VoltageLevel.BaseVoltage, but not with scl:Voltage, which is the targetontology relationship. Consequently, the alignment extractionmethod creates the SWRL alignment given in (8).

cim : VoltageLevelð?xÞ∧

cim : VoltageLevel:BaseVoltageð?x; ?yÞ-scl : Voltageð?x; ?yÞ ð8Þ

Data loss 3 – missing target relationship (inverse of sourcerelationship) occurs when:

� The instance i is mapped in Ds to ds and in Dt to dt.� The instance in is mapped in Ds to dsn and in Dt to dtn.� In Ds there is a relationship rs(ds,dsn).� In Dt there is an inverse relationship rt(dtn,dt).� In the output of the translation in is not related to i by the

relationship rt(in,i).


� From the standard class of ds and the relationship rs.� To the relationship rt, but changing the order of the arguments.


For example, in Fig. 13, the instance E1 is mapped to CIMVoltage-Level1 in Ds and to SCLVoltageLevel1 in Dt. Meanwhile, the instanceSE1 is mapped to CIMSubstation1 in Ds and to SCLSubstation1 inDt. Moreover, E1 is connected to SE1 with cim:VoltageLevel.

Substation, but not with scl:VoltageLevel, which is the targetontology relationship. Thus, the alignment extraction method createsthe SWRL alignment given in (9). It should be noted that in (9) theorder of the arguments in scl:VoltageLevel is the opposite to thatin cim:VoltageLevel.Substation.

cim : VoltageLevelð?xÞ∧

cim : VoltageLevel:Substationð?x; ?yÞ-scl:Voltageð?x; ?yÞ ð9Þ

Data loss 4 – missing target relationship (source relationshipby reference) occurs when:

� The instance i is mapped in Ds to ds and in Dt to dt.� The instance in is mapped in Ds to dsn and in Dt to dtn.� In Ds there is a relationship by reference between ds and dsn.� In Dt there is a relationship rt(dt,dtn).

� In the output of the translation i is not related to in by therelationship rt(i,in).


� From the standard class of ds and the relationship by referencebetween ds and dsn.

� To the relationship rt.

For example, in Fig. 14 the instance T1 is mapped to SCLTerminal1 inDs and to CIMTermina1 in Dt. Meanwhile, the instance L1 is mapped toSCLConnectivityNode1 in Ds and to CIMConnectivityNode1 in Dt. More-over, T1 and L1 are connected with an relationship by reference basedon the attributes scl:connectivitNode and scl:pathName; butthese instances are not connected with the object property cim:

Terminal.ConnectivityNode, which is the target ontology rela-tionship. Therefore, the alignment extraction method creates theSWRL alignment given in (10). It should be stressed that in (10), therelationship by reference between the terminal and the connectivitynode in SCL is represented by using the swrlb:string Equal

Ignore Case built-in (Horrocks et al., 2004).

scl : tTerminalð?xÞ∧scl : connectivityNodeð?x; ?vÞ∧scl : tConnectivityNodeð?yÞ∧scl : pathNameð?y; ?vnÞ∧swrlb : stringEqualIgnoreCaseð?v; ?vnÞ-cim : Terminal:ConnectivityNodeð?x; ?yÞ ð10Þ

Data losses other than the ones considered in this work mightoccur during the translations of the instance files. For instance, in theoutput of the translation (i.e., in graph I), there might not be arelationship by reference that exists in the target domain ontology(i.e., in graph Dt). With the aim of resolving this type of losses it wouldbe necessary to find complex alignments that can only be found byusing advanced data analysis and statistics methods. However, suchmethods require a numerous set of valid instance files (Euzenat andShvaiko, 2007), which are not always available in this context.

6.5. Iterative process

Sections 6.2–6.4 detailed how the Domain Expert proposes newalignments at each step of the iterative process. The functionwhose value indicates that the iterative process has ended should

MWBG

MWBG

E1Q1

L1 BB1 QA1

T3 T2 T1

SCLBay1

SCLBreaker1SCLConnectivityNode1

SCLTerminal2 SCLTerminal1

QA1

T2 T1

SCLBreaker1

SCLTerminal2 SCLTerminal1

Fig. 10. Hierarchical Mapping (HM).

Table 1Data losses considered in this study.

No. Name Description Example

Data losses detected by the Domain Expert1 Missing target class The data translator did not infer

the target ontology class of aninstance

see Fig. 11

2 Missing targetrelationship (equivalentto source relationship)

The data translator did not inferthe relationship that shouldconnect two instances in thetarget ontology. This targetrelationship is equivalent to thesource relationship that connectsthese instances in the sourceontology

see Fig. 12

3 Missing targetrelationship (inverse ofsource relationship)

The same as in 2, but in this casethe target relationship is aninverse property of the sourcerelationship

see Fig. 13

4 Missing targetrelationship (sourcerelationship byreference)

The same as in 2 and 3, but in thiscase the source relationship is arelationship by reference

see Fig. 14


measure the similarity between the output of the translations andthe target domain ontology. The measure of this similarity (sTOT) isobtained from the expression given in (11).

In (11): Rt(dt) is the set of all the relationships rt of the domainontology class dt, and s�rel(i,rt) is the similarity between arelationship rt of dt and the relationships of the instance i. Thevalue of s�rel(i,rt) is 1 if in the output of the translation there isan instance in related to i by the relationship rt(i,in). The value ofs�rel(i,rt) is 0.5 if the output of the translation does not includethe relationship rt(i,in) but it includes another relationship rs(i,in)

or rs(in,i). In any other case, s�rel(i,rt) is 0.

sTOT ¼∑iA I∑rtARtðdtÞs�gði; dtÞ �s�relði; rtÞ ð11Þ

The function sTOT is not unitary; this means that it can takevalues greater than 1. The unitary value sTOTm is obtained bydividing sTOT by the maximum sTOT, which is calculated by settingall the similarities s�g(i,dt) and s�rel(rt) at 1. The iterativeprocess stops when sTOTm does not change more than a thresholdε from the previous iteration. In this study, ε was set at 0.01.

SCLBreaker1

Standard Class:scl:tConductingEquipment

Fixed Value Attributes:scl:type = “CBR”

Attributes:scl:name

QA1

Classes:cim:Breaker

Attributes:cim:IdentifiedObject.name = “QA1”scl:name = “QA1”

CIMBreaker1

Standard Class:cim:Breaker

Attributes:cim:IdentifiedObject.name

??

Fig. 11. Data loss 1 – missing target class.

cim:VoltageLevel.BaseVoltage

cim:VoltageLevel.BaseVoltage

scl:Voltage

E1Classes:

cim:VoltageLevelscl:tVoltageLevel

Attributes:cim:IdentifiedObject.name = “E1”scl:name = “E1”

E1VClasses:

cim:Voltagescl:tVoltage

Attributes:cim:IdentifiedObject.name = “E1V”scl:name = “E1V”

CIMVoltageLevel1Standard Class:

cim:VoltageLevel


CIMVoltage1Standard Class:

cim:Voltage


SCLVoltageLevel1Standard Class:

scl:tVoltageLevel

Attributes:scl:name

SCLVoltage1Standard Class:

scl:tVoltage

Attributes:scl:name

??

Fig. 12. Data loss 2 – missing target relationship (equivalent to source relationship).


7. Evaluation

The first objective of the tests performed in this study was tocompare the proposed ontology matching system with the Agree-mentMaker (Cruz et al., 2007), which is currently one of the bestgeneric ontology matchers according to the Ontology AlignmentEvaluation Initiative (OAEI) competitions (Euzenat et al., 2011). Thesecond objective was to evaluate the MSC and the HM algorithmsdeveloped in this work. Thus, these algorithms were compared

against the main graph-based methods and the main mappingalgorithms included in the state of the art.

7.1. Case studies

In the case studies, the proposed system had to obtain thealignments for performing bi-directional translations betweenCIM and SCL of instance files representing five substation archi-tectures. Table 2 provides a brief description of these substation

cim:VoltageLevel.Substation scl:Voltage-

Level

CIMSubstation1Standard Class:

cim:Substation


CIMVoltageLevel1Standard Class:

cim:VoltageLevel


SCLSubstation1Standard Class:

scl:tSubstation

Attributes:scl:name

SE1Classes:

cim:Substationscl:tSubstation

Attributes:cim:IdentifiedObject.name = “SE1”scl:name = “SE1”

E1Classes:

cim:VoltageLevelscl:tVoltageLevel

Attributes:cim:IdentifiedObject.name = “E1”scl:name = “E1” SCLVoltageLevel1

Standard Class:scl:tVoltageLevel

Attributes:scl:name

cim:VoltageLevel.Substation

??

Fig. 13. Data loss 3 – missing target relationship (inverse of source relationship).

cim:Terminal.ConnectivityNode

SCLConnectivityNode1Standard Class:

scl:tConnectivityNode

Reference Attributes:scl:pathName

Attributes:scl:name

SCLTerminal1Standard Class:

scl:tTerminal

Reference Attributes:scl:connectivityNode

Attributes:scl:name

CIMConnectivityNode1Standard Class:

cim:ConnectivityNode


L1Classes:scl:tConnectivityNodecim:ConnectivityNode

Attributes:scl:name = “L1”scl:pathName=“SE1/E1/E1Q1/L1”cim:IdentifiedObject.name = “L1”

T1Classes:cim:Terminalscl:tTerminal

Attributes:scl:name = “E1”scl:connectivityNode =

“SE1/E1/Q1/L1”scl:substationName =“SE1”scl:voltageLevelName = “E1”scl:bayName = “Q1”scl:cNodeName = “L1”cim:IdentifiedObject.name = “E1”

CIMTerminal1Standard Class:

cim:Terminal

Attributes:scl:name

??

Fig. 14. Data loss 4 – missing target relationship (source relationship by reference).


architectures, which are: the simple substation (Simple) used in(Santodomingo et al., 2011), the radial substation (Radial)described in the IEC 61850-6 standard, and three representativesubstation architectures (Type_1, Type_2 and Type_3) defined bythe main Spanish electricity companies in (GSECS-61850, 2010).

7.2. Reference alignments

For each case study, a set of reference alignments has beencreated in this work by using Protégé 3.4.6 and Jena programmingframework (Jena, 2013). Each set of reference alignments includeall the required correspondences for performing bi-directionaltranslations between CIM and SCL for one of the substationarchitectures presented above. All the sets of reference alignmentshave been successfully validated in ESODAT, which proved able tomake valid data translations for all the substation architectures byimporting these reference alignments.

Table 3 describes the reference alignments. For the Simplearchitecture, all the required correspondences have beenexpressed in SWRL. However, this is not possible for the rest ofthe architectures. For the Radial architecture only 90.74% of thecorrespondences have been expressed in that language. The rest ofthe correspondences require the definition of new variables in theconsequent that have not been included in the antecedent of thetransformation rule. Therefore, these had to be represented in JenaRule Language by using the makeInstance built-in expression (Jena,2013). Given that the proposed ontology matching system onlycreates SWRL rules, in the Radial architecture it will be able toobtain at maximum 90.74% of all the required correspondences.

Table 3 also shows the number of simple equivalences that areincluded in the reference alignments. As discussed, most of theexisting ontology matchers are only able to obtain simple equiv-alences between the ontologies. Consequently, the ratio of equiv-alences determines the maximum percentage of alignments thatcan be obtained with generic ontology matchers. For instance, only29.63% of the alignments required for the Radial architecture canbe expressed as simple equivalences.

7.3. Results and discussion

This subsection provides a synthesis of the results obtained inthe tests. The complete results can be found in the website of the

first author.5 The tests were performed on a 3.16 GHz Intel CoreDuo CPU with 3.23 GB of RAM, running Windows XP.

7.3.1. Comparison with the AgreementMakerThe results obtained in the case studies with the ontology

matching system developed in this work were compared to thoseobtained with the AgreementMaker. In particular, the matcheremployed in the case studies was that used by the Agreement-Maker in the OAEI 2011 (Cruz et al., 2011). During the tests, it wasobserved that the maximum F-measures obtained with the Agree-mentMaker occurred with its mapping threshold set at 0.7. There-fore, this was the mapping threshold used in the AgreementMakerfor the case studies.

Table 4 shows the Recall (R), Precision (P) and F-measure (F)obtained in the case studies with: the schema-based subsystempresented in Section 5(S-based system), the complete ontologymatching system proposed in this work using the MSC graph-based method and the HM mapping algorithm (OM system), andthe AgreementMaker (AMaker). As can be seen, the completeontology matching system developed in this work improved theF-measures by almost 35% on average compared with the Agree-mentMaker. This is because the proposed system is able to findcomplex correspondences, improving the recall by 40% on average.Furthermore, due to the filter included in the instance-basedsubsystem, it also improves the precision by 22%.

As regards the performance measurements, the runtime withthe AgreementMaker and the schema-based subsystem was closeto 1 min; that is, as will be detailed in Section 7.3.2, almost fivetimes faster than with the complete system. However, given thatthe aligning between CIM and SCL ontologies is required forconfiguring power systems (not for real-time operations), theruntime with the matching system developed in this work isacceptable.

It should be emphasized that the comparison with the Agree-mentMaker is not a competition in equal terms, because due to theOAEI restrictions, the AgreementMaker does not use specificknowledge about the domain. In fact, this is the main contributionof the ontology matching system developed in this work: how toautomatically process deep domain knowledge in order to findsemantic correspondences that cannot be found with the bestgeneric ontology matchers.

With the aim of clarifying the scope and limitations of oursystem, the next three paragraphs will discuss the expected resultsof this system in a generic OAEI benchmark.

The OAEI benchmark test sets are built around a seed ontologyand many variations of it (Cuenca Brau et al., 2013). In that way,ontology matchers must align the seed ontology with all thevariations created in the benchmark. It is worth noting that thereference alignments in these benchmarks only include simplecorrespondences between pairs of entities. Hence, these tests do

Table 2Substation architectures used in the case studies.

Structure Bays Switches

Simple Two bays and one breaker 2 1Radial Radial 5 7Type_1 Double busbar – double busbar 26 95Type_2 One-and-a-half breaker to double busbar 43 151Type_3 H topology 21 71

Table 3Reference alignments.

Alignments SWRL rules (%) Equivalences (%)

Simple 26 100.00 46.15Radial 54 90.74 29.63Type_1 59 89.83 27.12Type_2 59 89.83 27.12Type_3 59 89.83 27.12

Table 4Compliance measurements with different systems.

S-based system OM system AMaker

R P F R P F R P F

Simple 0.39 0.71 0.50 0.89 0.96 0.93 0.39 0.53 0.44Radial 0.22 0.60 0.32 0.59 0.83 0.69 0.22 0.52 0.31Type_1 0.20 0.60 0.30 0.49 0.47 0.48 0.20 0.52 0.29Type_2 0.20 0.60 0.30 0.59 0.77 0.65 0.20 0.52 0.29Type_3 0.20 0.60 0.30 0.64 0.66 0.67 0.20 0.52 0.29Average 0.24 0.62 0.35 0.64 0.74 0.68 0.24 0.52 0.33

5 http://www.iit.upcomillas.es/santodomingo.


http://www.iit.upcomillas.es/santodomingo

not assess the ability of a system to find complex correspondences,which is the main feature of our system. Besides, in order tocompare ontology matchers in conditions that are equal for all ofthem, neither background knowledge nor instance files can beused in the OAEI benchmarks.

Taking the aforementioned restrictions into account, our ontol-ogy matching system would only be able to use 3 out of its8 ontology matching methods (see Fig. 5). All the matchingmethods of the instance-based subsystem require instances anddomain ontologies. For example, the graph-based methods areused in our system for mapping instances to domain ontologyclasses. Therefore, these methods would not be used by oursystem in the OAEI benchmark. As for the matching methods ofthe schema-based subsystem, only the string-based, language-based, and constraint-based methods would be employed in thebenchmark. As explained in Section 5, these methods are normallyenhanced in our system by using the electropedia domain termi-nology dictionary. However, without this specific backgroundknowledge of the electric power system domain, they are rathersimple methods that would obtain similar results to thoseobtained by the edna matcher, which is as a simple edit distancematcher that serves as a baseline in the OAEI evaluation cam-paigns. According to the OAEI report of 2011 (Euzenat et al. 2011),edna obtained an F-measure of 0.51 in one of the benchmarks,whereas the AgreementMaker obtained an F-measure of 0.71.These results show that when it is not possible to use backgroundknowledge and instance files, the best generic ontology matchersoutperform our system. Furthermore, generic ontology matchershave improved their performance during the last years. Thus, inthe benchmark used in the OAEI 2013 (Cuenca Brau et al., 2013),edna obtained an F-measure of 0.41, whereas the best ontologymatchers in those tests – YAMþþ (Ngo and Bellahsene, 2013) andCroMatcher (Gulić and Vrdoljak, 2013) – obtained F-measures thatwere close to 0.9.

In conclusion, we do not have developed a generic ontologymatcher that outperforms existing matchers in all cases. On thecontrary, we have defined a novel methodology to take advantageof background knowledge ontologies and instance files for findingcomplex semantic correspondences in a particular domain. Oursystem and methodology could be reused in other domains, butonly by importing domain ontologies and instance files of thatparticular domain. Otherwise, our system would not obtain asgood results as existing generic ontology matchers. This suggeststhat, in principle, it would be possible to enhance our system byadding novel techniques used by the best generic ontologymatchers. In particular, these techniques would improve theperformance of the domain-independent methods employed inthe schema-based subsystem in order to obtain better initialalignments. It should be noted, however, that this enhancementdoes not seem critical in the particular case of the CIM-SCLaligning. Thus, the tests carried out in this work show that theinitial CIM-SCL alignments found by the schema-based subsystemare similar to those found by the AgreementMaker (see Table 4).

7.3.2. Evaluation of MSC and HM algorithmsThe tests presented in this sub-section compared the results

obtained with the proposed system by using different similaritycomputations and different mapping algorithms in the instance-based subsystem. As explained in Section 6, the instance-basedsubsystem uses similarity computations and mapping algorithmsfor mapping instances to domain ontology classes. In that way,these tests assessed whether the innovative similarity computa-tion (MSC) and mapping algorithm (HM) developed in this workoutperformed the main solutions in the state of the art in thespecific task of mapping instances to domain ontology classes.

Table 5 shows the F-measures obtained with different similar-ity computations (rows of the table) and different mappingalgorithms (columns of the table). The notation used in Table 5for the similarity computations and the mapping algorithms wasexplained in Section 3.1. As can be seen, the best F-measure isobtained by combining the MSC with the HM. Furthermore, thiscombination obtains good results in terms of performance. Thus,whereas the average runtime in all the possible combinations is2544.77 s, the runtime in the MSC-HM combination is 303.196 s.

The explanation for these results is detailed as follows. MSC is asimple computation (hence, it obtains good performance measure-ments) that considers all the possible contributions of the asso-ciated nodes (sons, parents and siblings) in the propagation of thelocal similarities, which explains its good results in terms ofcompliance measurements. Regarding the mapping algorithm,HM only evaluates the similarities between nodes in the samehierarchical level. Therefore, it provides good results (in bothperformance and compliance measurements) when the equivalentconcepts in both ontologies are represented in the same hierarch-ical level, which occurs in the CIM and SCL aligning.

Additional tests were performed in order to draw broaderconclusions about the performance of MSC, HM, and, in general,about the performance of all the matching methods used in thisevaluation. In these additional tests, 10% of the hierarchicalrelationships included in the SCL domain ontology were trans-formed into normal relationships. In that way, we simulated thecase in which a number of equivalent concepts are represented indifferent hierarchical level in the ontologies to be aligned. Thetests showed that, after this modification in the hierarchies of thegraphs, the combinations SNM-MWBG and DIS-MWBG outper-formed the MSC-HM combination.

In summary, when the equivalent concepts in both ontologieswere in the same hierarchical levels, the combination MSC-HMimproved the F-measure by 1.5% compared with existing graph-based matching methods and mapping algorithms. However,when the hierarchies of the graphs were modified, other combi-nations obtained better compliance measurements. In any case, nomajor differences were found in the results obtained with differentgraph-based matching methods. Besides, none of them obtainedF-measures greater than 70%. Therefore, as will be explained inSection 8, other similarity computations different from graph-based matching methods should be used in future work toimprove the mapping of instances to domain ontology classes inthe instance-based subsystem.

8. Conclusions

In this contribution, an ontology matching system for futureenergy Smart Grids was presented. This system uses instance filesand imports deep domain knowledge from external domainontologies in order to obtain complex alignments that cannot befound with generic ontology matchers. Moreover, it employs

Table 5F-measures with different similarity computations and mapping algorithms.

PM SM MWBG HM

SF_simple 0.45 0.62 0.57 0.58SF_original 0.40 0.45 0.48 0.46SEFP 0.42 0.47 0.58 0.55DIS 0.39 0.37 0.65 0.62SSC 0.42 0.37 0.64 0.68SNM 0.46 0.58 0.67 0.61DSC 0.42 0.45 0.65 0.68MSC 0.43 0.50 0.66 0.68


innovative graph-based matching methods (DSC and MSC) andinnovative mapping algorithms (HM) that have been developed inthis work.

The evaluation of the proposed ontology matching system wasbased on the main interoperability issue within the scope of SmartGrids: the alignment of CIM and SCL ontologies. In the testsperformed in this study, the system had to obtain the requiredalignments to perform bi-directional translations between CIMand SCL files representing five substation architectures. In thesetests, the proposed system outperformed one of the best genericontology matchers according to the OAEI, the AgreementMaker.Furthermore, the MSC-HM combination obtained better resultsthan the main graph-based matching methods and the mainmapping algorithms included in the state of the art. It is worthnoting, however, that we have not developed a generic ontologymatching system. Thus, when background knowledge andinstance files cannot be used, the AgreementMaker and othergeneric matchers (such as, YAMþþ and CroMatcher) obtainbetter results than our system. Besides, the MSC-HM combinationoutperforms existing graph-based matching methods and map-ping algorithms only when the equivalent concepts in bothontologies are represented in the same hierarchical level, whichoccurs in the CIM and SCL aligning.

Even though the proposed system significantly improves theCIM-SCL aligning, it is not able to automatically find all thealignments. During the tests it was observed that the similaritycomputation was the bottle neck of the matching process; i.e., itdoes not seem possible to find more alignments from the similar-ity matrices created with the similarity computation. Therefore, inorder to improve the ontology matching system, another type ofsimilarity computation different from graph-based methodsshould be adopted for the system. On this topic, model-basedmethods, especially the methods based on Description Logics (DL),are a promising research direction. These methods have not beenconsidered in this study because, at the time of developing theontology matching system, they were not as mature as existinggraph-matching algorithms. Nevertheless, future work will ana-lyze in detail how DL-based methods can take advantage of thedefinitions included in the domain ontologies in order to comparethese domain ontologies precisely with the instance files importedin the system. Furthermore, additional case studies will bedesigned with the aim of proving the ability of the system toalign other standard ontologies in Smart Grids, or even in otherdomains.

Acknowledgment

The authors wish to thank Professor I.F. Cruz from the Uni-versity of Illinois at Chicago, for granting them permission to usethe AgreementMaker in this study. This work was supported inpart by the Ministry of Education of the Madrid Governmentunder Grant CPI/0349/2008.

References

Abanda, F.H., Tah, J.H.M., Duce, D., 2013. PV-TONS: A photovoltaic technologyontology system for the design of PV-systems. Eng. Appl. Artif. Intell. 26 (4),1399–1412.

Aleksovski, Z., 2008. Using background knowledge in ontology matching (Ph. D.dissertation). Vrije Universiteit, Amsterdam.

Bechhofer, S., van Harmelen, F., Hendler, J., Horrocks, I., McGuinness, D.L., Patel-Schneider, P.F., Stein, L.A., 2004. OWL Web Ontology Language Reference. W3CRecommendation. URL: ⟨http://www.w3.org/TR/owl-ref⟩.

Blomqvist, E., Öhgren, A., 2008. Constructing an enterprise ontology for anautomotive supplier. Eng. Appl. Artif. Intell. 21 (3), 386–397.

Cornell, 1999. A List of English Stop Words. Cornell University, Ithaca, New York (CSDept. URL: (ftp://ftp.cs.cornell.edu/pub/smart/english.stop).

Cruz, I.F., Antonelli, F.P., Stroe, C., 2009. Efficient selection of mappings andautomatic quality-driven combination of matching methods. In: Proceedingsof the 4th International Workshop on Ontology Matching, October 2009,pp. 49–51.

Cruz, I.F., Stroe, C., Caimi, F., Fabiani, A., Pesquita, C., Couto, F.M., Palmonari, M., 2011.Using AgreementMaker to align ontologies for OAEI 2011. In: Proceedings of the6th International Workshop on Ontology Matching, October 2011, vol. 814.

Cruz, I.F., Sunna, W., 2008. Structural alignment methods with applications togeospatial ontologies. Trans. GIS 12 (6), 683–711.

Cruz, I.F., Sunna, W., Makar, N., Bathala, S., 2007. A visual tool forontology alignmentto enable geospatial interoperability. J. Vis. Lang. Comput. 18 (3), 230–254.

Cuenca Brau, B. et al., 2013. Results of the Ontology Alignment Evaluation Initiative2013. In: Proceedings of the 8th International Workshop on Ontology Matching,pp. 61–100.

Euzenat, J. et al., 2011. Results of the ontology alignment evaluation initiative 2011.In: Proceedings of the 6th International Workshop on Ontology Matching,October 2011, vol. 814.

Euzenat, J., Shvaiko, P., 2007. Ontology Matching. Springer-Verlag, BerlinHeidelberg.

Euzenat, J., Valtchev, P., 2004. Similarity-based Ontology Alignment in OWL-Lite. In:Proceedings of the 16th European Conference on Artificial Intelligence (ECAI-04), pp. 333–338.

Falk, H., Saxton, T., 2010. Harmonizing the International Electrotechnical CommissionCommon Information Model (CIM) and 61850 Standards via a Unified Model: keyto achieve Smart Grid Interoperability Objectives. EPRI, Palo Alto, CA.

García, R., Gil, R., 2007. Facilitating Business Interoperability from the SemanticWeb. In: Abramowicz, W. (Ed.), Business Information Systems. Springer-Verlag,Berlin Heidelberg, pp. 220–232.

Gómez-Pérez, A., Fernandez-Lopez, M., Corcho, O., 2004. Ontological Engineeringwith examples from the areas of knowledge management, e-commerce and thesemantic web. Springer-Verlag, New York.

GSECS-61850, 2010. Minimum common specification for substation protection andcontrol equipment in accordance with the IEC 61850 standard. Group ofSpanish Electricity Companies for Studies on IEC 61850. URL: ⟨https://docs.google.com/viewer?a=v&pid=sites&srcid=ZGVmYXVsdGRvbWFpbnxlM2dyb3VwaWVjNjE4NTB8Z3g6MjAyMzMyNDI2YjZiMjcy⟩.

Gulić, M., Vrdoljak, B., 2013. CroMatcher – Results for OAEI 2013. In: Proceedings ofthe 8th International Workshop on Ontology Matching, October 2013, pp. 117–222.

Haslhofer, B., Klas, W., 2010. A survey of techniques for achieving metadatainteroperability. ACM Comput. Surv. 42 (2), 7:1–7:37.

Horrocks, I., Patel-Schneider, P.F., Boley, H., Tabet, S., Grosof, B., Dean, M., 2004.SWRL: A Semantic Web Rule Language Combining OWL and RuleML. WorldWide Web Consortium. National Research Council of Canada, Network Infer-ence, and Stanford University (URL). ⟨http://www.w3.org/Submission/SWRL⟩.

Jena, 2013. Jena Programming Framework for the Semantic Web – The generalpurpose rule engine. URL: ⟨http://jena.apache.org/documentation/inference/index.html#rules⟩.

Li, J., Tang, J., Li, Y., Luo, Q., 2009. RiMOM: a dynamic multistrategy ontologyalignment framework. IEEE Trans. Knowl. Data Eng. 21 (8), 1218–1232.

Marie, A., Gal, A. 2007. On the stable marriage of maximum weight royal couples.Presented at 6th International Workshop on Information Integration on theWeb (IIWeb), Vancouver, Canada, 23 July, 2007.

Mascardi, V., Locoro, A., Rosso, P., 2010. Automatic Ontology Matching via UpperOntologies: A Systematic Evaluation. IEEE Trans. Knowl. Data Eng. 22 (5),609–623.

Mc Namara, P., Negenborn, R., De Schutter, B., Lightbody, G., 2013. Weightoptimisation for iterative distributed model predictive control applied to powernetworks. Eng. Appl. Artif. Intell. 26 (1), 532–543.

Melnik, S., Garcia-Molina, H., Rahm, E., 2002. Similarity flooding: a versatile graphmatching algorithm and its application to schema matching. In: Proceedings ofthe 18th International Conference on Data Engineering, Feb–March 2002,pp. 117–128.

Morbach, J., Yang, A., Marquardt, W., 2007. OntoCAPE – a large-scale ontology forchemical process engineering. Eng. Appl. Artif. Intell. 20 (2), 147–161.

Munkres, J., 1957. Algorithms for the assignment and transportation problems.J. Soc. Ind. Appl. Math. 5 (1), 32–38.

Ngo, D., Bellahsene, Z., 2013. YAMþþ – Results for OAEI 2013. In: Proceedings ofthe 8th International Workshop on Ontology Matching, October 2013, pp. 211–218.

O'Leary, D.E., 1997. Impediments in the use of explicit ontologies for KBS develop-ment. Int. J. Hum. Comput. Stud. 46 (2), 327–337.

Porter, M.F., 1980. An algorithm for suffix stripping. Program 14 (3), 130–137.Rayudu, R.K., 2010. A knowledge-based architecture for distributed fault analysis in

power networks. Eng. Appl. Artif. Intell. 23 (4), 514–525.Rodríguez, M.A., Egenhofer, M.J., 2003. Determining semantic similarity among

entity classes from different ontologies. IEEE Trans. Knowl. Data Eng. 15 (2),442–456.

Sabou, M., d'Aquin, M., Motta, E., 2008. Exploring the semantic web as backgroundknowledge for ontology matching. J. Data Semant. 11 (2), 156–190.

Santodomingo, R., Rodríguez-Mondéjar, J.A., Sanz-Bobi, M.A., 2012. Using semanticweb resources to translate existing files between CIM and IEC 61850. IEEETrans. Power Syst. 27, 2047–2054.

Santodomingo, R., Rodríguez-Mondéjar, J.A., Sanz-Bobi, M.A., Rohjans, S., Uslar, M.,2011. Towards the automatic alignment of CIM and SCL ontologies. In:


http://refhub.elsevier.com/S0952-1976(14)00035-9/sbref1






















https://docs.google.com/viewer?a=v&pid=sites&srcid=ZGVmYXVsdGRvbWFpbnxlM2dyb3VwaWVjNjE4NTB8Z3g6MjAyMzMyNDI2YjZiMjcy







http://www.w3.org/Submission/SWRL

http://jena.apache.org/documentation/inference/index.html#rules

http://jena.apache.org/documentation/inference/index.html#rules


























Proceedings of the 2nd IEEE International Conference on Smart Grid Commu-nications (SmartGridComm), pp. 422–428.

Shvaiko, P., Euzenat, J., 2013. Ontology Matching: State of the Art and FutureChallenges. IEEE Trans. Knowl. Data Eng. 25 (1), 158–176.

Tartarus, 1980. Porter stemmer in Java. Tartarus people, projects and penitence.URL: ⟨http://tartarus.org/⟩.

Uslar, M., Rohjans, S., Schulte, S., Steinmetz, R., 2008. Building the Semantic Utility withStandards and Semantic Web Services. In: Meersman, R., Tari, Z., Herrero, P. (Eds.),On the Move to Meaningful Internet Systems: OTM 2008 Workshops. Springer-Verlag, Berlin Heidelberg, pp. 1026–1035.

Uslar, M., Specht, M., Rohjans, S., Trefke, J., Gonzalez, J.M., 2012. The CommonInformation Model CIM: IEC 61968/61970 and 62325 – A practical introductionto the CIM (Power Systems). Springer-Verlag, Berlin Heidelberg.

Winkler, W.E., 1999. The State of Record Linkage and Current Research Problems.Technical Report 99/04, Statistical Research Division, US Census Bureau.

Wriggers, P., Siplivaya, M., Joukova, I., Slivin, R., 2007. Intelligent support ofengineering analysis using ontology and case-based reasoning. Eng. Appl. Artif.Intell. 20 (5), 709–720.




http://tartarus.org/











ontology matching system for future energy smart grids

Documents