General asset management model in the context of an electric utility: Application to power transformers

Electric Power Systems Research 81 (2011) 2015–2037

Contents lists available at ScienceDirect

Electric Power Systems Research journal homepage: www.elsevier.com/locate/epsr

General asset management model in the context of an electric utility: Application to power transformers Juan L. Velasquez-Contreras a,∗ , Miguel A. Sanz-Bobi b , Samuel Galceran Arellano a a b

Polytechnic University of Catalonia, Department of Electrical Engineering, Diagonal 647, 08028 Barcelona, Spain Comillas Pontifical University, Institute for Research in Technology, IIT, Santa Cruz de Marcenado 26, 28015 Madrid, Spain

a r t i c l e

i n f o

Article history: Received 4 March 2009 Received in revised form 24 March 2011 Accepted 19 June 2011 Available online 19 July 2011 Keywords: Asset management Power transformers Detection Diagnosis Failure rate Maintenance

a b s t r a c t GAMMEU1 constitutes an integrated approach that covers the different elements related to the asset management of power transformers in the environment of a utility. GAMMEU harmonizes and interrelates all the relevant subsystems of the asset management that normally are studied as individual entities and not as a system. Concretely, GAMMEU consists of a platform for data integration, an intelligent system for detection and diagnosis of failures, a failure rate estimation model, a module of reliability analysis and an optimisation model for maintenance scheduling. In this work, a brief description of the elements of GAMMEU is presented and the implementation of the intelligent system for detection and diagnosis as well as the failure rate estimation model is exemplified using data of measurements performed in real power transformers. A robust anomaly detection module using prediction models based on artificial intelligence techniques was developed for top oil temperature monitoring and the use of decision trees as classifiers for the assessment of FRA2 measurements is also illustrated. For failure rate estimation, the use of a model based on hidden Markov chains is presented using data of dissolved gas analysis tests. The experience obtained from the implementation of part of the modules of GAMMEU using real data has demonstrated its feasibility. © 2011 Elsevier B.V. All rights reserved.

1. Introduction As a consequence of liberalization, investments in new transmission equipment have significantly declined over the past 15 years. Many transformers are working well beyond their intended life and are operating under increasing stress. As load is increasing, new generation, and economically motivated transmission flows push equipment beyond nameplate limits. As a result, business in the electrical sector has dramatically changed and for this reason, it is imperative to look for new opportunities and strategies for allowing electric utilities to survive these changes. In order to counterattack the undesirable consequences of the previously mentioned factors, utilities have looked for new methods and strategies that allow not only to achieve determined levels of reliability but also to do it in the most cost-effective manner. Among the different controllable elements that directly affect the network reliability, maintenance is the one of major relevance.

∗ Corresponding author. Tel.: +34 93 401 67 27; fax: +34 93 401 74 33. E-mail addresses: [email protected] (J.L. Velasquez-Contreras), [email protected] (M.A. Sanz-Bobi), [email protected] (S. Galceran Arellano). 1 GAMMEU: general asset management model for an electric utility. 2 Frequency response analysis. 0378-7796/$ – see front matter © 2011 Elsevier B.V. All rights reserved. doi:10.1016/j.epsr.2011.06.007

Between maintenance and reliability there is a clear relationship. If the equipment is not maintained, the probability of failure occurrence will increase, while in the case that the equipment is well maintained, it will be lower, but of course, at higher maintenance costs. In this sense, equilibrium between reliability and maintenance expenditure is required. As indicated in [1], the present state-of-the-art in maintenance strategies offers new opportunities which are structured in at least three basic approaches for making decisions related to maintenance. These opportunities are: (1) condition-based maintenance (CBM); 2) reliability centred maintenance (RCM); and (3) optimisation techniques (asset management/Risk Management). Ref. [2] presents an interesting work that shows the experience in Germany regarding the application of the RCM-strategy. By determining both condition and importance criteria, the strategy allows for determining which equipment has to be maintained first. It is worth mentioning the existence of a commercially available tool that works using this strategy [3]. Most of the authors who have written about this subject give positive opinions with regard to the implementation of the RCM strategy through condition and importance indices. Nevertheless, there are also some sceptical works, as the one indicated in [4], where it is stated that the implementation of RCM programs in this way represents a significant step in the direction of “getting the most out” of the equipment installed. However, the approach is still heuristic, and its applica-

2016

J.L. Velasquez-Contreras et al. / Electric Power Systems Research 81 (2011) 2015–2037

tion requires experience and judgement at every turn. Besides, it is stated that the RCM can take a long time before enough data are collected for making such judgements. For this reason, in [4], several mathematical models have been proposed to aid maintenance scheduling. Another tool is the system ADRES described in [5]. This system was developed as a decision making support tool for maintenance activities. As can be observed in this article, the decision support information is given in the form of a scoring system. On the other hand, information regarding reliability is an input data that should be provided instead of being calculated by the system. One major disadvantage detected in this work is that the system does not link the condition monitoring with the decision support tool. Another example is the maintenance management system SOFIA described in Ref. [6]. This tool uses condition models to estimate equipment condition in the future taking into account planned maintenance measures as well as the predicted environmental conditions. A case study for a long term maintenance program for one substation is presented including the life cycle cost analysis, which clearly indicates the benefits achieved by the new system. This system is basically a software implementation for maintenance scheduling based on CBM. The system predicts the condition and the costs of the maintenance activities and in this way, the maintenance works are planned. The main disadvantage of this tool is that the importance or the risks associated to the equipment are not taken into consideration. Nevertheless, as stated in this paper, great benefit has been achieved thanks to the use of this tool. The work presented in [7] is a very complete and interesting contribution that inspired the conception of the GAMMEU model. The main particularities of this work is the development of a platform for data integration and the application of different methods for estimating failure rates from condition data, which as will be seen next, constitutes one of the key features of the GAMMEU model. The main disadvantage of this work is that the failure rate estimation of the power transformers is only based on DGA (dissolved gasses analysis) and there is a lack of anomaly detection systems as well as expert systems required for guiding engineers during diagnostic tasks. In contrast to the previously mentioned work, the intelligent system for predictive maintenance SIMAP presented in [8] describes a robust approach for scheduling the maintenance activities of the gearbox of a wind turbine according to the information obtained from an anomaly detection module, a health condition assessment module and a diagnostic expert system, all of them developed under artificial intelligence techniques. The disadvantages of this method are on one hand the lack of a methodology for transforming the condition data into failure rate, and on the other hand, the requirement of knowledge about the history of faults for the development of a health condition assessment module. Taking as base the identified necessities and the revision of the tools already developed, the challenge of developing a new approach for satisfying the detected necessities was identified. As a result, GAMMEU was conceptualised and it characterises itself for having the following elements: A platform for data integration. This is of great importance given that the condition assessment task demands information not only about the condition, but also about the operation, maintenance and nameplate data. Additionally, by means of such a platform, the information exchange among elements of the models is facilitated. This idea was partially obtained from [7]. System for the detection of anomalies and their diagnostic. The application of artificial intelligence techniques is proposed for allowing the detection of anomalies and assessment of diagnostic measurements. The work of Ref. [8] was a source of inspiration for the conception of this system.

Failure rates estimation from information about the condition. The most important reliability index, i.e., the failure rate, will be estimated by a failure rate estimation model. This idea was inspired by the works quoted in [7,9–11]. Here it is important to mention that there are different methods for estimating failure rates by modelling the maintenance process [12,13]. These models establish a link between maintenance and reliability, which is a necessity that utilities have been expressing for a long time. The possibility of modelling maintenance becomes important when the objective consists of evaluating different maintenance policies (corrective maintenance or time-based maintenance). However, when maintenance is based on condition, the necessity of comparing different maintenance policies, in terms of reliability, changes to the necessity of transforming the condition information into failure rates. Reliability analysis. As reliability becomes more important to both utilities and customers, reliability studies will become just as important as, or even more important than power flow studies. For this reason, the importance of introducing a module for reliability analysis as part of GAMMEU has been determined. Optimisation module for maintenance scheduling. This module offers the flexibility of optimising and scheduling maintenance decisions under different objective functions and under defined technical and economical restrictions and availability of resources. Next the content of each of the elements of GAMMEU is described. First a general description of GAMMEU is presented in a form of a block diagram in Section 2. The concept of the platform for data integration and its technological aspects are presented in Section 3. Then in Section 4, each of the elements of the intelligent system of detection and diagnosis (anomaly detection module, diagnostic module and condition assessment module) are presented. The development and implementation of the anomaly detection module is illustrated for top oil temperature monitoring using records of an on-line monitoring system installed in a 30 MVA transformer. While for the illustration of the diagnostic module, the methodology for the implementation of an intelligent system of assessment of frequency response analysis (FRA) is presented and its application on real transformers is exemplified through real case studies. The methodology for the implementation of the condition assessment module is also described in Section 4 and exemplified using real data of diagnostic measurements. Section 5 deals with the estimation of the failure rate of transformers based on the results of the condition assessment module presented in Section 4. As example, records of DGA measurements of a 40 MVA transformer that failed 34 months after its installation were used for illustrating the implementation of the failure rate estimation model. In Section 6, the failure rate estimated in Section 5 is proposed to be used for computing the mean time to failure (MMTF) data required for conducting reliability analysis. The reliability indices obtained from Section 6 are proposed to be used in Section 7 for the development of an optimisation model for maintenance scheduling. After the conceptualisation of GAMMEU and its elements, for a feasible real implementation, in Section 8 GAMMEU is adapted to the environment of a utility, where the typical information technology tools used by the utilities (ERPS, MMS, etc.) are represented and linked to the elements of GAMMEU. 2. General description of GAMMEU GAMMEU is based on the features mentioned in Section 1. As can be observed in Fig. 1, the model consists of the following elements: • • • • •

Information technology platform for data integration Intelligent system for detection and diagnosis Failure rate estimation model Reliability analysis Optimisation model for maintenance scheduling.

J.L. Velasquez-Contreras et al. / Electric Power Systems Research 81 (2011) 2015–2037

2017

Fig. 1. Block diagram of GAMMEU.

3. Information technology platform for data integration Electric utilities have different IT (information technology) systems, belonging to different departments that have been developed over several years. These systems have different user interfaces and their own databases, and may be running on different operating systems. Personnel working in a given department increasingly expect information from other systems to be readily available. In most cases, exchanging information among different systems is a cumbersome process. Some utilities have undertaken large integration projects where a great deal of effort is spent on creating a bespoke integration, integrating each application point-to-point. Arguably, this is not the optimal way of solving the integration challenge, and some of these projects fail. If they do succeed, the integration solution is still hard to maintain, and replacing or adding a system requires changes in all integrated systems [14]. For the above reasons mentioned, the necessity of incorporating a platform for data integration was formulated. The IT platform plays a key role in the GAMMEU model, since it facilitates the possible integration of the data coming from different sources in just a single platform. The role of this IT platform in maintenance management of utilities is essentially the establishment of communication among data. From the intelligent system for detection and diagnosis point of view, the data to be integrated are: records of condition (tests and monitored parameters), records of operation; records of maintenance and technical data of the transformers (nameplate data). Once all the before mentioned data are collected and integrated, the intelligent system of detection and diagnosis will be able to carry out its tasks, that is, anomaly detection, interpretation of diagnostic tests and condition assessment. The role of the IT platform is also the establishment of communication links required among the other elements of the GAMMEU model, i.e., failure rate estimation model, reliability analysis and the optimisation model for maintenance scheduling. In [15], two approaches for data integration are presented: Data Warehousing and Database Federation. In the Data Warehousing approach, data from heterogeneous information sources is gathered, mapped to a common structure and stored in a central location. Periodic updating is required to ensure that the information contained in the warehouse is up-to-date with the contents of the individual sources. However, the data replication/updating

process can be quite expensive in the case of large information repositories. Also, this approach relies on a single common ontology for all users which is specified as part of the warehouse design. As a result, this system tends to be less flexible. On the other hand, in the case of Database Federation, the information needed to answer a query is directly gathered from the data sources in response to the posted query. Hence, the results are updated with respect to the contents of the data sources at the time the query is posted. This approach is being more readily adapted to applications where users are able to impose their own ontologies and specify queries using the various concepts defined in these ontologies. Because of this, sometimes Database Federation is preferred. There are also other integration concepts such as the Aspect Object technology presented in [14]. With the Aspect Objects technology, each physical object is represented by an Aspect Object, and the functionality related to the object is modelled as aspects of the object. One aspect might be a document with operating procedures; another could be a live video of the physical object. A number of standard aspect types exist for standard documents, applications and web links. Other examples that give evidence of technological advances in the field of data integration and its application are presented in [16–18]. 4. Intelligent system for detection and diagnosis “Future intelligent power grids consisting of compact equipment with integrated diagnostic intelligence providing self-healing and self-diagnosing facilities will become the norm where high performance materials and knowledge about the degradation processes will be vital [19]”. In harmony with this prediction, the intelligent system for detection of anomalies and diagnosis has been conceived for carrying out the following functions: • Anomaly detection (by means of the anomaly detection module) • Diagnosis (by means of the diagnostic module) • Condition assessment (by means of the condition assessment module). In Fig. 2 these modules, as well as the relationships among them are illustrated. Records of on-line measurements are sent to AIOMS where the task of anomaly detection is carried out. If AIOMS detects any anomaly, diagnostic tests have to be scheduled for diagnosing

2018

J.L. Velasquez-Contreras et al. / Electric Power Systems Research 81 (2011) 2015–2037

Fig. 2. Modules of the intelligent system for detection and diagnosis.

such anomaly. The execution of the diagnostic tests is scheduled by the optimisation model for maintenance and scheduling. The results of these tests are stored in the database of off-line condition records where the results of the historical diagnostic tests carried out on the transformers are stored. The interpretation of the diagnostic tests is done by the diagnostic module where for each diagnostic method an automatic assessment of the results is provided. For the interpretation of the diagnostic tests, information about the nameplate data of the transformers, records of operation (e.g. number and through faults) and records of maintenance (e.g. previous repairs) are also very useful. The outcomes of the assessments done by individual diagnostic methods are sent to a condition assessment module where these are combined and an objective judgement on the condition of the transformer as a condition index is done. The penetration of intelligent systems based on artificial intelligence (AI) and data mining techniques into power systems is a reality. Even in the field of monitoring and diagnosis there are many evidences that confirm this fact [20–23]. These modern techniques have a huge potential for the development of each of the functions of the intelligent system for detection and diagnosis. Even with the help of AI techniques, the tasks of detection and diagnosis used to be complicated due to the complexity and diversity of the tasks. For experts this often implies simultaneously tackling different types of knowledge (inaccurate, incorrect or redundant) from different data sources that require being processed using different reasoning mechanisms [24]. In the field of computer science, multi-agent systems have gained popularity over the last years due to their capabilities of solving detection and diagnostic problems in complex domains. As an example of this, in [25] the implementation of a multi-agent system for condition monitoring of two in-service transmission transformers in the UK is presented. Multi-agent systems are composed of multiple interacting computing elements, known as agents [26]. These agents are computer systems with two important capabilities. First, they are, at least to some extent, capable of autonomous action. Second, they are capable of interacting with other agents in a social collaboration.

Considering that in the detection and diagnosis of failures in power transformers different techniques are used, it is suitable to use a multi-agent approach for the development the modules of intelligent system for detection and diagnosis. 4.1. Anomaly detection module The detection of anomalies is carried out by means of an automatic and intelligent on-line monitoring system (AIOMS). As shown in Fig. 3, AIOMS is based on a multi-agent system by means of which a robust detection of anomalies can be archived. The agents shown in Fig. 3 allow detection of thermal, electrical, degradation and mechanical failure modes in the active part of power transformers as well as mechanical and thermal failure modes in the tap changer and degradation and electrical failure modes in bushings. On-line detection of mechanical anomalies in power transformers such as axial and radial deformations in windings is a challenging and difficult task. Some research efforts have been devoted to on-line monitoring of vibrations [31,32] and transfer function [33,34] has been tested on real applications to power transformers. On-line monitoring of vibrations is a promising technique that in the near future could become consolidated. This motivated the definition of Agent 9 (vibration monitoring) within the framework of AIOMS. AIOMS has as its main objective the identification of symptoms of failure modes in each component of power transformers taking into account ambient and operational conditions. The input data for this module is the condition information (mainly from on-line measured data) to be obtained through the platform of data integration. There are different methods for detecting anomalies from on-line data: limit checking, trend analysis, pattern recognition and model-based monitoring. Limit checking consists of comparing actual measurements with configured limit values. A notification (alarm) is generated in the case that the limit values are exceeded. Trend analysis consists of discerning whether the level of a measured variable has increased or decreased over time, and if it has, how quickly or slowly the increase or decrease has occurred. Pattern recognition is the act of taking in raw data and taking an action

J.L. Velasquez-Contreras et al. / Electric Power Systems Research 81 (2011) 2015–2037

2019

Fig. 3. Block diagram of AIOMS.

based on the category of the pattern. Pattern recognition aims to classify data (patterns) based either on a priori knowledge or on statistical information extracted from the patterns. The patterns to be classified are usually groups of measurements. Model-based monitoring has as its aim to predict the behaviour of key variables able to respond to the evolution of failure modes. The prediction can be done either by the first principle method (white box models) or by artificial intelligence (AI) models (black box models). The first principle method (FPM) uses mathematical models based on process physics to predict the behaviour of the desired variables, while AI models uses the explicit expert knowledge or the knowledge hidden in examples saved in databases for training algorithms able to learn about the dynamics of a system and its environment. Once the algorithms have learnt, these are able to predict the behaviour of the variables of interest. The use of model-based monitoring has gained a lot of attention over the last years due to its superior and more reliable performance. In summary, by using any of the previously mentioned methods of anomaly detection, each of the agents of AIOMS can be developed. Taking into consideration that some failure modes such as thermal failures could be detected by more than one agent (e.g. a thermal failure will increase the oil temperatures and generate gasses), it makes sense to combine the outcome of each agent in an expert system in order to configure a coherent diagnosis. The expert system is in charge of embedding human knowledge into the system for determining whether a failure mode is present or not in the transformer (judgement). Some intelligence can also be added for allowing the expert system to distinguish between failure modes in components and failures in the sensors themselves. In the case that the presence of a failure mode is detected, the expert system is also in charge of recommending a list of diagnostic tests to be carried out for diagnosing the detected failure mode. As a result, the user is notified (either via mail, SMS, etc.) about the event occurred and a downloadable report is generated.

Details on the implementation of AIOMS are illustrated in this work only for Agent 1 (top oil temperature monitoring, TOT). For the implementation of other agents, similar methodology can be followed. The first step for the implementation consists on defining the monitoring method to be used. The first principle method (physical thermal models) has been formulated for TOT prediction [35,36]. This method has as disadvantage that several thermal rated parameters of the transformers are required to the model, but normally these values are unknown. Due to the lack of information on these rated thermal parameters of the transformer, in previous research works conducted by other authors two possible solutions to this problem have been explored. One solution could be to use AI techniques for identifying the parameters of the model [27–30,37–39] and other solution could be to use AI models (black box models) [40–42]. The experience with black box models based on artificial neural networks (ANN) is encouraging. According to [42] the prediction with ANN is similar to the results obtained with semi-physical models. In [42] it was also found that ANN has a better performance against rapid ambient temperature changes. Based on the positive experiences of other authors with ANN for modelling the top oil temperature monitored, this approach was chosen for the implementation of the Agent 1 of AIOMS. The second very important step is the selection of the input variables to train the ANN. The typical input variables used for the prediction of the top oil temperature in an ANN model are: load current, ambient temperature and operating states of the cooling system (pumps and fans). These are the same variables required by the semi-physical models to estimate the top oil temperature. As a general rule, the input variables are chosen as those variables explaining the output variable (explicative variables). It is advisable to avoid unnecessary input variable to the model in order to make it robust. A classical way for selecting the input variables is by means of sensitivity analysis (scatter plots, correlation coefficients, transfer function models, principal components analysis, etc.). It is logic

2020

J.L. Velasquez-Contreras et al. / Electric Power Systems Research 81 (2011) 2015–2037

Fig. 5. Topology of the ANN model for TOT prediction.

Fig. 4. View of the transformer and details of installation of a Pt100 sensor at the top of the tank.

to think that the TOT depends on the load current, on the ambient temperature and on the operating states of the cooling system, where these are exogenous variables to the transformer and therefore these were pre-selected as input variables. But considering that the effect of changes in these exogenous variables could affect the TOT with a certain time constant, the final selection of the variables was done after analysing on-line data taken from a 30 MVA power transformer. Details on the design of this system are presented in [1,43]. Different variables are measured on-line in this transformer. Fig. 4 shows a view of the transformer and details of the installation of a Pt100 sensor at top of the tank. From the analysis of the data it was found that, as expected, there is a time constant delaying the relationship between changes in the load current and in the TOT. The best correlation between the current and the TOT was obtained under a time constant of 45 min (during summer months). It was also found that changes in the ambient temperature have an almost immediate effect on the TOT. Since the cooling system of the transformer in which the monitoring system was installed is normally out of service, in this case the operating condition of the cooling system has not an important impact as input variable. Based on these findings, the variables chosen as input to predict the TOT in the time instant “t” are: load current with a time delay of 45 min (t − 45 min) and the ambient temperature in “t”. A feed-forward back-propagation ANN was used. The ANN was trained using the Matlab Neural Network Toolbox. The architecture of the ANN is shown in Fig. 5. It has two input variables (load current at “t − 45 min” and the ambient temperature at “t”). A hidden layer consisting of 8 neurons was determined to be suitable. The neurons of the hidden layers use the tangent sigmoid as activation function. The output layer corresponds to the variable to be predicted, that is, the TOT. It consists of a single layer with a linear activation function. The network was trained using data recorded during July and August of 2008 (a total of 5602 samples). After training the ANN, its prediction capabilities were tested by means of a validation data

Fig. 6. Measured versus predicted TOT using data of June of 2009.

set collected in June 2009 (2868 samples). As shown in Fig. 6, the ANN was able to predict the behaviour of the TOT satisfactorily. The performance of the ANN was assessed by statistical analysis of the prediction error. The error has a mean value of 0.108 ◦ C, a standard deviation of 2.459 ◦ C, a minimum of −6.787 ◦ C and a maximum of 7.793 ◦ C both in respect to de value expected. After training the ANN, control limits were established using the method EWMA (exponentially weighted moving average chart) as illustrated in Fig. 7. This chart was generated from the error data of the validation set giving a weight of 0.9 and multiples of 3 of the standard deviation. According to the results, under normal behaviour operation of the transformer, the deviations between the model and the actual measurements shall be between +6.78 ◦ C and −6.57 ◦ C. However, from the anomaly detection point of view, only the upper limit is of interest. Any measurements point out of the upper limit is given as input to the expert system in charge of issuing a judgement. For the judgement of thermal anomalies, the expert system of AIOMS embeds human knowledge in the form of inter-relation diagrams under which the outcome of two or more agents are interrelated with each other for confirming the outcome of each single agent on one side, and for allowing a basic localization of the thermal anomaly on the other side. The creation of such inter-relation

J.L. Velasquez-Contreras et al. / Electric Power Systems Research 81 (2011) 2015–2037

Fig. 7. Establishment of control limits for TOT monitoring using the EWMA chart.

diagrams is one of the benefits of working under a multi-agent concept, in which the performance of one agent can be inter-related to the performance of other ones for achieving the same objective. The inter-relation diagram used by AIOMS for explaining out-ofexpected values of the top oil temperature measurement is shown in Fig. 8. As shown in Fig. 8, if the variable TOT is out of the control limits, the expert system first of all checks if the out-of-expected behaviour is temporal or sustained in a period of time, where temporal duration corresponds to a short period of time not greater than 15 min. In this way, if the TOT is out of control for more than 15 min, a sustained behaviour is identified and the expert system checks if other related agents, in this case, the bottom oil temperature and tap changer temperature monitoring agents, are also out of the control limits. According to expert knowledge, a thermal failure in the active part should be observed by both TOT and BOT monitoring and depending on the severity of the failure and its location

2021

it could also be observed by the TCT monitoring. As shown in Fig. 8, if TOT is out of control limits the expert system checks if BOT and TCT are also out of limits. If BOT is simultaneously out of limits, it on one side confirms the presence of a thermal problem and it also indicates that probably the thermal failure is located in the medium zone of the winding. But if BOT is out of limits with a certain time delay, it is logical to think that the thermal problem is located in the upper zone of the winding and in this case measures are to be taken. This situation could be confirmed by infrared inspections, but in any case, inspections are to be scheduled for diagnosing the cause of the thermal anomaly. Additionally, due to the heat transfer process that occurs in the transformer, any thermal problem in the upper zone of the winding shall anyhow affect the behaviour of the bottom oil temperature. For this reason, if TOT is out of control and BOT does not get out of control after certain time, the presence of a thermal failure is questionable and in this case a check of the performance of the sensors is recommended. Thermal monitoring of the tap changer oil temperature can also be inter-related to the TOT monitoring. In case that a thermal anomaly exists in the active part, it is expected to observe its effect in the tap changer compartment, but with a delay. A simultaneous out-of-control state of the variables TOT and TCT is not very probable and the occurrence of this situation is an indication of simultaneous thermal failures in the windings and in the tap changer what is also not very likely. Similar inter-relation diagrams can be developed for explaining the out of behaviour expected for the measurements of bottom oil temperature and tap changer oil temperature. Similar methodology can be used for the development of the other agents. However it is convenient to highlight that not always the implementation of a model-based continuous monitoring is possible. For certain variables such as gasses dissolved in oil, the development of models is much more challenging and in some cases, not feasible. For this type of variables, threshold checking or trending are more suitable methods.

Fig. 8. Inter-relation diagram for detection of thermal anomalies by TOT monitoring.

2022

J.L. Velasquez-Contreras et al. / Electric Power Systems Research 81 (2011) 2015–2037

4.2. Diagnostic module The objective of the diagnostic module is to confirm the judgement provided by the AIOMS and identify possible failure modes that might be in evolution in the transformer before a major or catastrophic failure occurs. This module is also expected to be able to determine the nature of the anomaly and if possible its location. It consists of two blocks. One block of agents based on traditional methods of diagnosis and other block of agents based on modern methods of diagnosis [44]. Traditional agents are conceived for an automatic and intelligent assessment of: dissolved gasses in oil (AIADGA), physical–chemical and electrical analysis (AIAPCE), furan analysis and degree of polymerization (AIAFDP), traditional electrical tests (AIATET) and infrared inspections (AIAIRI). AIATET covers the assessment of all the typical electrical tests performed on-site in power transformers: windings resistance measurements, voltage ratio, excitation current, short-circuit impedance, frequency response of stray losses, dissipation factor and insulation resistance. The following modern diagnostic agents have been proposed: automatic and intelligent assessment of frequency response analysis (AIAFRA), dielectric response (AIADRM), frequency response of stray losses (AIAFRS) and partial discharges (AIAPD). Considering that shutdowns of power transformer can be very costly for the utilities, the diagnostic module will consider at first traditional diagnostic agents that do not required disconnecting the transformer (AIADGA, AIAPCE, AIAFDP and AIAIRI). In case that the anomalies detected by AIOMS are confirmed by these diagnostic agents, the disconnection of the transformer is justified and by means of AIATET and modern diagnostic agents (AIAFRA, AIADRM, and AIAPD) a comprehensive and reliable condition assessment of the transformer can be achieved. The implementation of the diagnostic module consists of different agents as described in Section 4.2. Regarding the agents for assessment of traditional diagnostic methods (AIADGA, AIAPCE, AIAFDP, AIATET and AIAIRI), fortunately along the years a considerable experience has been collected and at the present, several international standards and institutions provide guidelines for the interpretation of the results [45–47]. The challenge in this field is the achievement of a consensus among different assessment criteria. The implementation of these agents is based on the collection of different assessment criteria and applying artificial intelligence techniques such as fuzzy logic for the uncertainty treatment of the assessment based on individual standards. Additionally as a complement, the knowledge hidden in the databases of measurements of diagnosed transformers is also proposed to be used for the development of these agents. Regarding the agents for assessment of modern diagnostic methods (AIAFRA, AIADRM, AIAFRS, and AIAPD) the situation is much more complicated due to the lack of standards for interpretation of results. Particularly, the interpretation of FRA measurements is a cumbersome task that at the present can only be carried out by human experts. Given the added value of FRA measurements in the detection of mechanical deformations and their importance for condition assessment of power transformers, a research work based on real FRA measurements with a support of a manufacturer of FRA instruments was initiated for the development of AIAFRA [48]. Next a short overview to the implementation of AIAFRA is presented as an example about how GAMMEU can integrate this kind of methods. 4.3. AIAFRA Although the frequency response analysis (FRA) method is well established as a powerful tool for detection and diagnosis of mechanical and electrical failure modes in the active part of power

transformers, the interpretation of results still requires the analysis and expertise of a human expert. AIAFRA means “Automatic and intelligent assessment of FRA measurements”. The word “automatic” means that the results will be issued without the participation of human experts. However, it is worth to mention that the role of the human expert in this field can only be minimized, but it is irreplaceable. The word “intelligent” means that AIAFRA shall posses the intelligence and knowledge of human experts. These two simple words suggest very big challenges: first, an algorithmic representation of the acting conducts of the human experts (philosophies, procedures, etc.) and second, a representation of the knowledge and experience of the human experts are required. FRA is a comparative measurement method, this means that the results of an actual test are compared with a reference test, also called fingerprint. The framework of AIAFRA was conceived for supporting the comparison methods that can be used for the assessment of the results. Three comparison methods can be used to assess FRA results: a. Time-based comparison (TBC). Current FRA results are compared to previous results of the same transformer. b. Phase-based comparison (PBC). FRA results of one phase are compared to the results of the other phases of the same transformer. c. Construction-based comparison (CBC). FRA results of one transformer are compared to another one of the same design. In the literature a distinction between sister transformers and twin transformers is given. Twin transformers are defined as identically designed and identically assembled transformers, while sister transformers have the same maker, the same specifications and the same customer but with different manufacturing years, which can be very likely related to different designs since the economic and quality boundaries oblige transformer manufacturers to optimise their designs and production processes constantly. For the assessment, human experts first of all carry out a TBC and in case they find suspicious deviations, a PBC is normally used as a complement to the TBC. If the reference test required for TBC is not available, human experts try to perform the assessment by the PBC method. According to the collected experience from a database of FRA measurements performed in more than 500 power transformers, it was found that in many cases a PBC allows carrying out a reliable assessment if the FRA plots of the different phases keep good symmetry among each other. It was also found that normally HV (high voltage) windings used to be quite symmetrical. Nevertheless, the FRA plots of MV (medium voltage) and LV (low voltage) windings tend to be more unsymmetrical. In those cases in which due to asymmetries is not possible to provide a reliable assessment, human experts look for the FRA results measured in a twin/sister transformer for carrying out a CBC. These are the principles that define the acting conducts of human experts. Now that the acting conducts of human experts is somewhat defined by the comparison methods before described, the next aspect to be considered within the AIAFRA framework is the representation of knowledge and experience of human experts. From this, it is expected to interpret what a deviation between two FRA plots means. It should be mentioned that even for human experts with several years of experience in the field of FRA, it is not always easy to provide a 100% reliable diagnosis because there are uncertainties that cannot be avoided. But in any case, with the present state of knowledge and based on previous experiences it was thought that is possible represent in some way the available knowledge, where available knowledge is understood as the knowledge hidden in databases of FRA measurements carried out in transformers with a known diagnosis. According to this, in the

J.L. Velasquez-Contreras et al. / Electric Power Systems Research 81 (2011) 2015–2037

2023

Fig. 9. User and development workflows of AIAFRA.

framework of AIAFRA a database of FRA measurements of diagnosed transformers is to be considered. A diagram showing the flow of the data together with the different elements of the AIAFRA framework is presented in Fig. 9. A distinction has to be established between the user flow and the development flow. The user’s flow is represented with dashed rows and starts at the “user” who has to upload the data to be assessed by means of any kind of user interface (web server, etc.) and ends downloading a report provided by the expert system. The development’s flow, shown with solid rows, starts with noise detection and removal of the FRA data (block 1 in Fig. 9). Afterwards, an automatic identification of frequency sub-bands is done (block 2 in Fig. 9) and a third step the FRA data is transformed into information by means of extraction of indicators (block 3 in Fig. 9). The extracted indictors are used for training transforming information into knowledge and for that machine learning algorithms are used. The extracted knowledge represent the knowledge base (block 5 in Fig. 9) necessary for development of the expert system of AIAFRA (block 6 in Fig. 9) that is in charge of interacting with the user. As can be appreciated in Fig. 9, as part of the development workflow the role of a human expert is also included. A human expert shall be in charge of confirming the validity of the automatic assessment provided to the users and on the other side in charge of deciding which of the new cases have to be added to the FRA database in order to enhance the diagnosis capabilities of AIAFRA. Depending on the diversity of the previous experiences stored in the FRA database, it could happen that some new cases are not correctly classified. In this case, the human expert detects this erroneous assessment and assigns a correct diagnosis to the case and sends it to the FRA database. Afterwards, the switch S1 is closed, what means that the knowledge extraction process is repeated and the new extracted knowledge is sent to the knowledge base by closing the switch S2. After finishing this updating

process, the switches S1 and S2 are opened and the system is ready to assess new cases, but this time with better classification capabilities because there is a wider knowledge and with better quality. Next the steps to be followed for the development of AIAFRA are illustrated and subsequently its application is demonstrated by real case studies. As shown in Fig. 9, the first step of the development consists of noise detection and removal. The noise that could be present in the plots is automatically detected and removed. Then, transformer-specific sub-band identification is carried out. The proposed sub-band structure consists of five sub-bands: two low frequency sub-bands (LF1 and LF2), where the magnetizing inductance (Lm) and the parallel capacitances (Cg) dominate the response, one medium frequency sub-band (MF), where the interaction among windings dominates, and two high frequency sub-bands, HF1 and HF2, where the winding structure and the grounding and internal leads dominate the response, respectively. Just for illustration purposes, the frequency sub-bands of a 200 kV winding are illustrated in Fig. 10. For the automatic identification of these sub-bands, the algorithms presented in [49] have been used. The extraction of indicators (feature extraction) can be carried out by means of different algorithms. Statistical indicators such as standard deviations and correlation coefficients are preferred due to their simplicity. However any other kind of indicators, which are able to measure the deviation among curves can be used. In this work, the extraction of indicators using correlation coefficients is illustrated. The correlation coefficient (CC) of two time series of n elements, X {x1 , x2 , x3 , . . ., xn } and Y {y1 , y2 , y3 , . . ., yn } is defined by Eq. (1). CC =

n  x=i



xi yi

n x2 x=i i

n

y2 x=i i

(1)

2024

J.L. Velasquez-Contreras et al. / Electric Power Systems Research 81 (2011) 2015–2037 Table 1 Failure mode classes.

Fig. 10. Illustration of the proposed frequency sub-band structure.

If for each frequency sub-band a set of elements X and Y is considered for both magnitude and phase of the FRA response and the sub-band correlation coefficients can be computed. For each case ten CCs have been extracted: five magnitude CCs (CCMLF1, CCMLF2, CCMMF, CCMHF1, CCMHF2) and five phase CCs (CCPLF1, CCPLF2, CCPMF, CCPHF1, CCPHF2). Once the data is transformed into information in the form of indicators, the hidden knowledge is extracted. Some algorithms used in the fields of data mining and machine learning [50,51] have the capabilities of learning complex patterns from examples included in databases. The AIAFRA framework includes a block called “Knowledge Extraction” which is based on machine learning algorithms for extracting the required knowledge for automatic assessment of new cases. In this paper the application of decision trees as machine learning algorithm is shown. Specifically, the decision tree algorithm C4.5 implemented in the open source machine learning software Weka was used. The input to this algorithm is a training set consisting of “instances”, which represent the previous experiences of human experts stored in the database. Each instance is described by a set of “attributes” belonging to one “class”. The attributes are the extracted indicators. In this particular case, the ten sub-band CCs

Failure mode classes

Abbreviation

No. instances

Healthy winding under the same remanence condition (HESRC) Healthy winding under different remanence condition (HEDRC) Short-circuit between turns (EFST) Mechanical deformation (MEDE)

A

38

B

78

C E

32 34

mentioned above are the attributes, while the classes are normally chosen as the failure modes to be diagnosed. After reviewing and analyzing each of the FRA measurements in the database in which a time-based comparison was possible, it was found that the majority of the measurements correspond to healthy transformers, as expected. For some failure modes such as open-circuit (EFOC) and (EFCR), among others, the number of instances in the database was very small. Taking into account that classes having very few instances are not representative of the real patterns of the deviations of such classes, it was decided not to include them within the training set. As a result only the classes listed in Table 1 were considered as part of the training set. Because short-circuit between turns influence the low frequency range of the FRA response and mechanical deformations influence the FRA response at high frequencies, it was decided to develop two independent decision trees for the diagnosis these failure modes. The decision tree in charge of the short-circuit diagnosis between turns (class C) is shown in Fig. 11, while the decision tree in charge of diagnosis of mechanical deformations (class E) is shown in Fig. 12. As can be appreciated, the decision tree in charge of diagnosing short-circuit between turns also is able to classify the class B (healthy windings under different remanent conditions). This is important because in some cases the deviations at low frequencies caused by remanent magnetism might be misinterpreted as short-circuits between turns. Fig. 13 shows an example of a diagnosis provided by AIAFRA in a 30 MVA transformer. The correlation coefficients needed for the assessment are as follows: CCMLF1 = 0.999; CCMHF1 = 0.927; CCMHF2 = 0.788; CCPLF1 = 0.999; CCPLF2 = 0.999; CCPMF = 0.942; CCPHF1 = 0.490; CCPHF2 = 0.328. According to the decision tree in charge of diagnosis short-circuit between turns, the transformer is healthy (class A). However, the transformer has a mechanical defor-

Fig. 11. Decision tree for diagnosis of short-circuit between turns.

J.L. Velasquez-Contreras et al. / Electric Power Systems Research 81 (2011) 2015–2037

2025

Fig. 12. Decision tree for diagnosis of mechanical deformations.

CCPMF = 0.841; CCPHF1 = 0.998; CCPHF2 = 0.985. According to the decision trees, the transformer has short-circuited turns (class C), but no mechanical deformations. The two examples presented above illustrated the validity of AIAFRA for automatic assessment of FRA results in real power transformers. 4.4. Condition assessment module

Fig. 13. Example of diagnosis of mechanical deformations with AIAFRA.

mation (class E), according to the classification of the decision tree in charge of mechanical deformations. Fig. 14 shows an additional example of assessment of a 40 MVA transformer provided by AIAFRA. The correlation coefficients needed for the assessment are as follows: CCMLF1 = 0.998; CCMHF1 = 0.998; CCMHF2 = 0.883; CCPLF1 = 0.996; CCPLF2 = 0.278;

Fig. 14. Example of diagnosis of short-circuit between turns with AIAFRA.

The condition assessment tasks of GAMMEU are carried out by the agent for automatic and intelligent condition assessment (AICA). AICA has as objective to provide an overall assessment of the transformer in the form of a condition index “CI”. In the literature, some methods have been proposed [52–54], however, these methods do not consider the integration of on-line monitoring, traditional diagnostic methods and modern diagnostic methods in a single approach. AICA is founded on a discrete representation of the condition deterioration process of a power transformer by means of a multistate condition model (MSC model). The MSC model represents the deterioration of the condition during the life span of the transformer in five stages (new, normal, defective, faulty and failed), as suggested in [55]. Additionally, the condition at each stage is represented by different states. For each of the states a numerical value was assigned, where this value corresponds to the condition index (CI) of the transformer. Fig. 15 shows the proposed MSC model. As can be observed, the model has 5 stages and 11 states. According to the topology of the model a new transformer having a CI = 10 is said to be new and a transformer having a CI = 0 or CI = 1 is said to be failed. The challenging task of the condition assessment performed by AICA consists in mapping the results of monitoring and diagnostic measurements into the MSC model shown in Fig. 15. This is a complex task due to the diversity of failure modes that could occur in power transformers and due to the difficulties in combining the interpretation of results obtained from different monitoring and diagnostic methods. AICA addresses these difficulties by means of a matrix of detection and diagnosis of failure modes (DEDIFA). DEDIFA on one side establishes the inter-relation among monitoring and diagnostic methods and on the other side establishes the possible stages and states of each failure mode in the MSC model. As example, in this paper only the part of the matrix DEDIFA corresponding to diagnostic methods is illustrated in Fig. 16. Table 2 shows the description of the abbreviations used in Fig. 16 for the diagnostic methods. DEDIFA also represents the effectiveness of each diagnostic method for the diagnosis of each failure mode by means of a numerical value, here called certainty. For those diagnostic methods having a high effectiveness, a certainty value of 0.9 was assigned,

2026

J.L. Velasquez-Contreras et al. / Electric Power Systems Research 81 (2011) 2015–2037

Fig. 15. Topology of the multi-state condition model of AICA.

while for diagnostic methods having medium and low effectiveness a certainty of 0.6 and 0.4 was assigned, respectively. As before discussed, the failure modes in DEDIFA as well as the diagnostic methods and the stages and states of the MSC model are inter-related with each other. But, this is not enough for an automatic condition assessment. As part of the condition assessment a systematic methodology to reach a consensus about the overall condition of transformers is needed. For that, AICA uses a

new method introduced paper called “weighted-class consensus method (WCCM)” as a methodology allowing the combination of different monitoring and diagnostic methods. This method is based on multi-agent systems under which each monitoring and diagnostic method is treated as an agent. The WCCM is applied for each failure mode in an independent way. For that, first of all, all diagnostic methods able to trace the failure mode “i” are identified from the matrix DEDIFA. For example,

Fig. 16. Matrix DEDIFA.

J.L. Velasquez-Contreras et al. / Electric Power Systems Research 81 (2011) 2015–2037 Table 2 Diagnostic methods and their abbreviations in DEDIFA. DGA PCEA COND MORS MPED MPIS MPKF FUR DPO TTR EXCU MABA DCWR DF/PF INRE/POI CGRO SCI IRI FRA FRA-SC FRSL FRCL FRLI FRDF DFTU DRM PD

Dissolved gas analysis Physical–chemical and electrical analysis of oil Oil conductivity Relative saturation of moisture in oil Moisture in paper based on equilibrium diagrams Moisture in paper based on sorption isotherms Moisture in paper based on Karl Fischer titration Furan analysis Degree of polymerization Transformer turns ratio Exciting current Magnetic balance test DC winding resistance Dissipation factor/power factor Insulation resistance/polarization index Core grounding test Short-circuit impedance Infra red inspections Frequency response analysis Frequency response analysis short-circuit test Frequency response of stray losses Frequency response of core losses Frequency response of leakage inductance Frequency response of dissipation factor Dissipation factor tip-up test Dielectric response methods Partial discharges

for the case of the failure mode short-circuit between turns, only the following diagnostic methods would participate in the consensus: DGA, TTR, EXCU, MABA, DCWR and FRA. Fig. 17 shows the block diagram of the WCCM method. As can be appreciated, each diagnostic method is treated as an agent. In this context, each agent characterizes itself for having an input (data of monitoring and diagnostic measurements), intelligence for automatic assessment and an output (stage–state position in the MSC model). The outcome of the individual diagnosis provided by each agent is represented by means of the multi-agent class voting matrix (CVM). Subsequently, the combination of the diagnosis of all agents participating in the diagnosis of the failure mode “i” is

2027

achieved by means of the multi-agent class consensus process, as next described. The multi-agent class consensus process is carried out by means of a weighted voting process in which the votes of the agents, represented by the class voting matrixes (CVM) are weighted by certainty factors represented by the certainty matrixes (CM). These matrixes are presented by Eqs. (2) and (3).

⎛

a11

⎜ . CVM = ⎜ .. ⎝

am1

⎛

b11

⎜ . CM = ⎜ .. ⎝

bm1

···

a1n

..

.. .

.

···

b1n

..

.. .

···

⎟ ⎟ ⎠

(2)

amn

··· .

⎞

⎞ ⎟ ⎟ ⎠

(3)

bmn

In a multi-agent environment consisting of n agents and m classes to be classified, the CVM has a size mxn, where each element aij = 1, with 1 ≤ i ≤ m and 1 ≤ i ≤ n, if the agent i votes for the class j and aij = 0 otherwise. The classes to be classified correspond in this case to the position (stage, state) occupied by a transformer in the MSC model. The certainty matrix consists of certainty factors representing how accurate, truthful or reliable the diagnosis provided by each monitoring and diagnostic method for each specific failure mode is. The certainty matrix also has a size mxn, where each element bij equals the certainty factors of the matrix DEDIFA shown in Fig. 16. The outcome of the agents is unified by the multi-agent class consensus. For that a numerical value called brute utility (BU) is used as a measure of the weighted outcome of each agent and can be calculated by Eq. (4). In view that higher the number of agents voting in favor to the same class, higher the level of confidence of the assessment, the utility before defined is treated as a brute utility (BU) and a net utility (NU) is obtained by multiplying BU by a coincidence factor (CF) as indicated by Eq. (5), where CFi = 0.5 if

Fig. 17. Block diagram of the WCCM method for combination of diagnostic methods.

2028

J.L. Velasquez-Contreras et al. / Electric Power Systems Research 81 (2011) 2015–2037

Table 3 Illustration of the multi-agent consensus process. Class

CVM

CM

Class consensus

ME 1

ME 2

ME j

ME n

ME 1

ME 2

ME j

ME n

No. votes in favor

BU

CF

NU

Stage 1, state 1 Stage 2, state 1

a11 a21

a12 a22

a1j a2j

a1n a2n

b11 b21

b12 b22

b1j b2j

b1n b2n

K K

BU1 BU2

CF1 CF2

NU1 NU2

C1 C2

Stage i, state k

ai1

ai2

aij

ain

bi1

bi2

bij

bin

K

BUi

CFi

NUi

Ci

Stage m, state n

am1

am2

amj

amn

bm1

bm2

bmj

bmn

K

BUm

CFm

NUm

Cm

only one agent votes in favor to the class i, CFi = 0.8 if two agents votes and CFi = 1 if 3 or more agents vote in favor. BUi =

n 

CVMik × CM1,k

(4)

k=1

NUi = BUi × CF

(5)

As a result, from the consensus the class with the highest NU is chosen as winner class and a new certainty factor (C) is assigned to this class, where the certainty factor corresponds to the coincidence factors before mentioned. In case that two or more classes simultaneously fulfill the condition of highest NU, the winner class corresponding to the worst position in the MSC model is chosen as winner. The agent class consensus process is summarized in Table 3. As can be appreciated, the classes of the consensus process correspond to the stage/state positions of the MSC model, while the agents correspond to the diagnostic methods (ME 1, ME2, . . ., MEn) participating in the consensus process of the failure mode “i”. After determining the number of votes in favor (K) to each class and calculating the BU and NU values, the winner class is determined and a certainty factor C is associated to it. Fig. 18 shows the workflow of AICA. Fist users upload the data of the diagnostic methods to be used for performing a condition assessment. AICA identifies the failure modes from the available diagnostic measurements that can be traced by such diagnostic methods (traceable failure modes, FM1, . . ., FMn). Afterwards, for each of the traceable failure modes AICA performs a condition con-

Winner class

C

sensus using the WCCM method. From the outcome of the condition consensus of each failure mode, an overall condition consensus for the whole transformer is carried out and a user report is generated. Next, the workflow of AICA is illustrated by means of a real case of condition assessment of a 100 MVA, 230/115 kV transformer. The goal of the condition assessment was to assess only the condition from the point of view of degradation. For that, the diagnostic methods listed in Table 4 were applied. According to the workflow of AICA, as the first step, the traceable failure modes are identified. The failure modes that can be traced by the diagnostic method listed in Table 4 are shown in Table 5. The next step of the workflow is the performance of a condition consensus for each traceable failure mode using the WCCM method. For simplification purposes, in this paper only the condition consensus for the failure mode “degradation due to water in oil” is illustrated. According to Table 5, the following diagnostic methods are able to trace the failure mode “degradation due to water in oil”: MORS, COND, CR, DS, ODF and WCO. As illustrated in Fig. 17, these diagnostic methods are treated by AICA as agents that are characterized for having intelligence for automatic assessment. Such intelligence could be knowledge extracted from databases using machine learning algorithms (as illustrated in Section 4.3) or could be interpretation limits recommended by standards or institutions. As example, Table 6 shows the intelligence used for the agent in charge of assessing the measurements of the method MORS. In this case the intelligence of the agent is based on an adaptation of the limits recommended by the standard IEEE Std 62 [47] to the MSC

Fig. 18. Block diagram of the workflow of AICA.

J.L. Velasquez-Contreras et al. / Electric Power Systems Research 81 (2011) 2015–2037 Table 4 Diagnostic methods used for condition assessment. Method

Abbreviations

Value

Relative saturation of moisture in oil (%) @20 ◦ C Moisture in paper using equilibrium diagrams (%) Frequency domain spectroscopy (%) Oil conductivity @20 ◦ C (pS/m) obtained from FDS 2-Furaldehyde (ppm) Colour comparator number Dielectric strength (kV) Oil dissipation factor @20 ◦ C Interfacial tension (dynes/cm) Water content in oil (ppm)

MORS MPED DRM COND FUR CR DS ODF IFT WCO

4.26 1.4 1.5 0.130 0.73 2 91 0.002 35 5

Table 5 Traceable failure modes. Methods used for condition assessment

Traceable failure modes

Relative saturation of moisture in oil (%) Moisture in paper using equilibrium diagrams (%) Frequency domain spectroscopy (%)

Degradation due to water in oil Degradation due to water in paper Degradation due to water in paper Degradation due to water in oil, degradation due to aging by-products in oil Degradation due to aging by-products in paper Degradation due to aging by-products in oil Degradation due to water in oil, degradation due to aging by-products in oil Degradation due to water in oil, degradation due to aging by-products in oil Degradation due to aging by-products in oil Degradation due to water in oil

Oil conductivity @20 ◦ C (pS/m) obtained from FDS 2-Furaldehyde (ppm) Colour comparator number Dielectric strength (kV)

Oil dissipation factor

Interfacial tension (dynes/cm) Water content in oil (ppm)

model. In a similar way, the intelligence of the other agents was defined. Table 7 shows the results of the condition consensus for the failure mode “degradation due to water in oil”. As can be observed, from this consensus it was determined that from the point of view of the agents (diagnostic methods) able to trace this failure mode, the condition of the transformer in the MSC model, represented by the winner class, corresponds to the stage “new”, as this is the class having the highest net utility (NU). After performing the condition consensus for the rest of the traceable failure modes, AICA performs an overall condition consensus. For that, the stage and the state of the failure mode having the worst stage–state position in the MSC model are chosen as winner. Table 8 summarizes the outcome of condition consensus of each of the traceable failure modes. As can be appreciated, the fail-

2029

ure mode having the worst stage–state position in the MSC model corresponds to the failure mode “degradation due to aging byproducts in paper”. In this manner, the overall condition consensus indicates that the condition of the transformer occupies the stage “defective” in the state 1. At the same time, according to the MSC model, this position corresponds to a condition index of 6. The condition assessment report is illustrated in Table 9. The reports shows the goal of the assessment, in this case, degradation, and indicates that from the point of view of degradation, the 100 MVA transformer has an overall condition index of 6 with a certainty of 0.5. A short text describing the assessment is provided. The report provides also a detailed overview of the assessment of each failure modes with its corresponding certainties, so that users are able to see how reliable the assessment of each failure mode is.

5. Failure rate estimation model “A statistical approach based on published reliability data can be useful as a first step in estimating maintenance and investment budgets. However, each transformer should be considered individually so that the user can decide whether it should be maintained, relocated, retrofitted or replaced [56]”. As response to this necessity, the automatic and intelligent failure rate estimation model (AIFRE) of GAMMEU was conceived. AIAFRE has the goal to transform the CI determined by the condition assessment module (AICA) into a failure rate, which is the main required value for conducting reliability analysis. For carrying out this function only few methodologies has been found in literatures. Records of failures spanning multiple components over an extended time period is one of the most commonly used methods for calculating failure probabilities. These can be classified into two categories: parametric and non-parametric estimation. Since power transformers are crucial and expensive equipments in transmission systems, they usually are well maintained and consequently have very high reliability. So in reality transformer failures are relatively rare, and it is difficult to obtain statistically significant failure data. Due to the difficulties of having enough failure data, recently some methods based on the evolution of the deterioration process have been proposed. These methods are: hidden Markov models (HMM) [57,58], Bayesian Analysis [9] and one method based on failure rate modelling using equipment inspection data [10]. In this work, a modification of the method presented in [57] is proposed to be used for transforming the CI into a failure rate. The objective of the HMM is to determine the transition probabilities among the stages that represent the condition of the transformer during its life time. For that, first of all the number of non-observable and observable states of the HMM have to be defined. A representation consisting of 5 non-observable states (new, normal, defective, faulty and failed) is chosen in harmony with the stages of the deterioration process described in Section

Table 6 IEEE 62 limits for relative moisture saturation in oil adapted to the MSC model. IEEE 62

Settings for the condition model

Limits (%)

Condition

Stage

State

Limits (%)

0–5 6–20

Dry Moderate to wet

New Normal

21–30

Wet

Defective

>30

Extremely wet

1 1 2 3 1 2 3 1 2

0 ≤ MSO ≤ 5 5 < MSO ≤ 10 10 < MSO ≤ 15 15 < MSO ≤ 20 20 < MSO ≤ 25 25 < MSO ≤ 30 30 < MSO ≤ 40 40 < MSO ≤ 50 MSO > 50

Faulty

2030

J.L. Velasquez-Contreras et al. / Electric Power Systems Research 81 (2011) 2015–2037

Table 7 Condition consensus for the failure mode “degradation due to water in oil”. Class

CVM

CM

Stage

State

MORS

DS

New Normal

1 1 2 3 1 2 3 1 2

1

1

Defective

Faulty

ODF

Class consensus

WCO

MORS

DS

ODF

WCO

No. votes in favor

BU

CF

NU

Winner class

1

0.9 0.9 0.9 0.9 0.9 0.9 0.9 0.9 0.9

0.6 0.6 0.6 0.6 0.6 0.6 0.6 0.6 0.6

0.6 0.6 0.6 0.6 0.6 0.6 0.6 0.6 0.6

0.9 0.9 0.9 0.9 0.9 0.9 0.9 0.9 0.9

3 0 1 0 0 0 0 0 0

2.4 0 0.6 0 0 0 0 0 0

1

2.4

New

0.5

0.3

1

0

Table 8 Overall condition consensus. Failure mode

Stage–state position

Certainty (C)

Winner overall condition

Degradation due to water in oil Degradation due to water in paper Degradation due to aging by-products in oil Degradation due to aging by-products in paper

New Normal, 1 Normal, 1 Defective, 1

1 0.8 0.8 0.5

Defective, 1

4.3. At the same time, a total of 11 observable states (O1, . . ., O11) corresponding to the value of the CI are defined as shown in Fig. 19. A HMM is characterized with the following parameters: a. Markov transition matrix: state transition probabilities A = {aij }, aij = p(qt+1 = j|qt = i), 1 ≤ i ≤ N, where qt denotes the current state. b. Probability of getting an observation with a symbol under specific state B = {bj(k)}, bj(k) = p{ot = vk |qt = j}, 1 ≤ j ≤ N, 1 ≤ k ≤ M, where ot denotes the current observation. c. Initial state distribution ˘ = {i}, where i = p{q1 = i}, 1 ≤ i ≤ N.

This is a learning problem, in which the HMM parameters are adjusted so that the CIs of a fleet of transformers are represented by the model in the sense of maximum likelihood, which means to get the optimal parameter  = {A,B,П}, by maximizing the likelihood of observation Ltot = p(O|). The Baum Welch algorithm was used for performing this task. After determining the HMM transition intensities, the transition probability matrix (P) can be obtained by the Eq. (6).

⎛1−a

12

⎜ ⎝

P=⎜

Table 9 User report. Transformer data Serial number: 12345678, rated power: 100 MVA, rated voltages 230/115 kV, age: 36 years Goal of the assessment Assessment of degradation Overall condition index = 6 The transformer was assessed as defective in state 1 due to degradation of the paper insulation. Certainty of the assessment = 0.5 Detailed overview of the assessment for each failure mode Description

MSC position

Certainty

Recommendations

Degradation of insulation due to water in oil Degradation of insulation due to water in paper Degradation of insulation due to temperature Degradation of insulation due to aging by-products in oil Degradation due to corrosive sulphur Degradation of insulation due to aging by-products in paper

New

1

None

Normal, 1

0.8

None

Degradation of insulation due to discharges

No assessed

No assessed

Normal, 1

0 a23 1 − a34 0 0

0 0 a34 1 − a45 0

⎞

0 0 ⎟ 0 ⎟ ⎠ a45 1

(6)

The state probability vector gives the probability that a component is in any particular deterioration level at a given time, and is denoted by: p(hT) = [p1(hT) · p2(hT) · p3(hT) · p4(hT) · p5(hT)], where h = 1, 2, 3, . . ., and T is the time increment. If at time t = 0, the component resides in the stage new (stage 1), then the initial state probability vector is P(0) = [10,000]. The probability of finding the component in any deterioration stage at the time hT is then given by P(hT) = P(0) × Ph , where the last number of each probability vector p(hT) corresponds to the probability that the transformer is in the stage failed (stage 5) before time hT, or the CDF (cumulative density function) of failure. The failure rate, which is the instantaneous probability of the component to fail during the period of [(h + 1)T, hT], given the con-

Perform tests able to track this FM 0.8

No assessed Defective, 1

0 0 0 0

a12 1 − a13 0 0 0

0.5

None

Perform tests able to track this FM Trend analysis for assessing the evolution and confirm with other methods Perform tests able to track this FM

Table 10 Transition times and CI indices of a 40 MVA transformer during its useful life. Stage

TDCG

State

CI

Time (weeks)

1 2

555 603 723 853 1007 1308 1768 2044 3057 4069 5082

1 1 2 3 1 2 3 1 2 1 2

11 10 9 8 7 6 5 4 3 2 1

12 30 17 18 25 5 3 11 11 11 9

3

4 5

J.L. Velasquez-Contreras et al. / Electric Power Systems Research 81 (2011) 2015–2037

2031

Fig. 19. Topology of the HMM adapted to the deterioration process of power transformer.

Then the failure rate was computed as the probability of finding the transformer in the stage 5 (failed). This was done by taking the last number of the probability vector P(hT), i.e., the column 5 at different times hT and by applying the Eq. (7). The failure rate obtained using this procedure with the data shown in Table 9 is shown in Fig. 20. The implementation of this model is also suitable for a fleet of transformer. In this case, the average transition time from one state to another can be used for training the HMM. 6. Reliability analysis

Fig. 20. Behaviour of the failure rate estimated using DGA data of a 40 MVA transformer.

dition that it survives to time hT can be calculated according to the Eq. (7). Pr (hT ≤ x ≤ (h + 1)T |x > hT ) =

P((h + 1)T ) − P(hT ) 1 − P(hT )

(7)

For illustrating the implementation of the failure rate estimation model, the deterioration process of a 40 MVA transformer which had a failure after 34 months of operation is used. The deterioration process was characterized using data coming from dissolved gas analysis tests performed in the transformer and from its installation till the failure. Specifically the total dissolved combustible gasses (TDCG) method was used and the methodology described in Section 4.4 was applied. For each state of each stage, the corresponding CI is given and the time that the transformer remained in each of the states is computed as shown in Table 10. These times are used for generating the sequence of the state distribution ˘ under which the HHM was trained using the function hmmtrain implemented in the Matlab Statistics Toolbox. The model was trained giving initial values to the matrixes A and B and the state distribution ˘. The resultant matrix P is:

⎛ 0.9762 0.0238 0 0 0.9714 0.0286 0 ⎜ 0 P=⎜ 0 0.9615 0.0385 ⎝ 0 0 0

0 0

0 0

0.9756 0

⎞

0 0 ⎟ ⎟ 0 ⎠ 0.0244 1

In general, the assessment of reliability indices for a complete or partial power system network is the availability assessment of the network to provide the connected customers with electrical energy, critical aspect to keep a good quality of service. The typical input data for conducting reliability studies is the mean time to failure (MTTF) and the mean time to repair (MTTR). Usually, for reliability studies, typical values published several years ago, such as the values presented in the references [59–61], are used. Considering that the design, working and aging conditions of power transformers have changed in the last years, it has been found that the use of typical failure statistics is always limiting and is potentially misleading [62]. In this work, the MMTF is obtained from the failure rate estimation proposed in Section 5. As the repair times have not dramatically changed, the typical statistical values of MTTR are proposed to be used. As a result, load point indices and system indices can be obtained from the reliability analysis. The load point indices are: expected failure rate [interruptions/year], annual outage time [hours/year], average outage duration [hours/interruption and average energy not supplied [kWh/year]. The system indices are: SAIFI [interruption/year, customer], SAIDI [hours/year, customer], CAIDI [hour/interruption], and AENS [kWh/year, customer]. At the same time, these indices are proposed to be used for evaluating the risks behind different maintenance actions not only on the reliability of transformers, but on the reliability of the network be means of the optimisation model for maintenance scheduling presented in Section 7. For the computation of the reliability indices there are commercially available solutions (e.g. DIgSILENT, NEPLAN, ETAP, PSS/E, etc.). For this reason, the development of a new methodology in the framework of GAMMEU is not necessary.

2032

J.L. Velasquez-Contreras et al. / Electric Power Systems Research 81 (2011) 2015–2037

Fig. 21. Optimisation model for maintenance scheduling.

7. Optimisation model for maintenance scheduling “Maximum asset performance is one of the major goals for electric power distribution system operators (DSOs). To reach this goal minimal life cycle cost and maintenance optimisation become crucial while meeting demands from customers and regulators [63]”. The optimisation model of GAMMEU was defined for reaching this goal. The model carries out several functions as can be observed in Fig. 21. The model has as inputs the outcome of the condition assessment carried out by the Intelligent System of Detection and Diagnosis. From such condition assessment a list of possible maintenance actions are automatically generated by the maintenance actions module. For each of the generated maintenance actions, an estimation of the associated cost (D 1, D 2, D i . . ., D N) and effects of the actions is carried out by the cost estimation module and by the effect estimation module, respectively. The effect estimation module calculates to each action a new condition index (CI1, CI2, CIi, . . ., CIN) by following the methodology presented in Section 4.4. Subsequently, these calculated CIs are sent to the failure rate estimation model where the failure rate is re-calculated. Then a reliability analysis is carried out and based on these results an estimation of the reduction of the risks to be expected thanks to the execution of the maintenance action “i” is done. The estimated costs and risk reductions are sent to the optimisation module which is in charge of scheduling the maintenance actions to be done on each transformer taking into account the availability of resources (e.g. human resources, facilities, spare parts/units), technical restrictions (like for example minimum reliability levels) and economical restrictions (e.g. available budge for maintenance). The scheduling can be done in short-term, medium term and long term, depending on the definition of the objective function to be optimised. The effect estimation module can be based on extracted knowledge from databases. For example, if a database of moisture

measurements performed before and after oil drying are available, these data can be used for estimation of the effects of the maintenance action “oil drying”. As example, the results of moisture measurements carried out before and after drying a 130 MVA, 220/110/45 kV Transformer manufactured in 1967 can be mentioned. The measured moisture in oil before drying was of 24 ppm (1.15 kg H2 O by weight), what corresponds to approximately 3% water content in paper (144 kg H2 O by weight). After drying, the estimated water content in paper was of 120 kg (2.5%). This kind of data can be used for estimation of the effects of maintenance actions on the condition of the transformer. In view of the complexity of implementing optimisation procedures of maintenance scheduling integrating all sub-systems show in Fig. 21, as it should be, alternative methods for maintenance scheduling based on prioritization diagrams that make use of condition and importance indexes (CI and II) have been proposed by other authors [2]. Using asset prioritization diagrams on one side is possible to classify assets from the point of view of maintenance into three groups: assets requiring maintenance, assets that have to be replaced or repaired and assets in which only inspections (monitoring and diagnostic measurements) have to be performed under regular basis. On the other side, prioritization diagrams allow establishing a ranking representing the priority under which the execution of maintenance actions should be scheduled. Before describing how the asset prioritization diagram works, a clear definition of condition and importance indexes is provided. The condition index (CI) is a numerical value representing the overall condition of an asset. Usually the CI corresponds to the CI value generated by the condition assessment methodology presented in Section 4.4. However, additional aspects related to the condition can be used for complementing the CI obtained from condition assessment. Examples of additional aspects considered for the condition index are:

J.L. Velasquez-Contreras et al. / Electric Power Systems Research 81 (2011) 2015–2037

2033

Fig. 22. Decision support system based on asset prioritization diagrams.

• Age of transformers • Serviceability, comprising, e.g. the availability of spare parts and skilled staff • Ambient conditions • Loading, as related to thermal stresses are highly relevant for many aging processes • General operational experience with similar assets • Availability of spare parts for aged equipment. For all these criteria a quantitative evaluation scheme and weight factors have to be agreed so that at the end, one single condition index can be calculated. The importance index (II) is a numerical value reflecting the technical, economic and social consequences of a potential failure. In other words, this index is a result of a risk assessment. It is based in criteria such as: • • • • • • • • •

Reliability indexes Voltage level and location of the equipment Cost of investing in replacement asset Service staff response times Topology of the network Security of supply Importance of the customer Degree of redundancy Economic consequences of a failure (e.g. penalties, loss of revenues from energy trading) • Environmental impact of a failure.

Similar as for the CI, for all these criteria a quantitative evaluation scheme and weight factors have to be agreed so that at the end, one single importance index can be calculated. The prioritization diagram is a Cartesian diagram having the importance index in the abscissa axe and the condition index in the ordinate axe, as shown in Fig. 22. After determining numerical values for the importance and condition indexes, the point represented by the ordered pair (II, CI) is plotted in the diagram. The diagram has three zones which are determined by the threshold values CIR and CIM . The first zone (green zone) is the zone between the highest value of the condition index (i.e. 10) and CIM , the second zone (yellow zone) corresponds to the zone between CIM and CIR and the third zone (red zone) has its contour lines between CIR and the lowest value of the condition index (i.e. 0).

The values to be chosen for the thresholds CIM and CIR are settings to be provided according to the experience. Depending on the zone to which an ordered pair (II, CI) belongs, the maintenance action to be scheduled is determined according to the classification rules shown to the right side of Fig. 22. These rules indicate that for those transformers whose ordered pairs (II, CI) belong to the first zone only regular inspections have to be scheduled, while for those transformers whose ordered pairs (II, CI) belong to the second zone, maintenance actions (for example oil drying) have to be scheduled. In a similar way, transformers having an ordered pair belonging to the third zone have to be replaced or repaired. Now that the way how the selection of the type of maintenance action corresponding to each asset was explained, the next question is about how to prioritize among assets. It could happen that after plotting the condition and importance indexes of a fleet of 200 transformers of a utility it is determined that in 85 transformers maintenance action have to be performed in 2012. But considering the hypothetical case that the maintenance budget for 2012 can only cover the maintenance costs for 60 transformers in 2012, the question is about how to prioritize in a systematic way the 60 transformers to be maintained in such way that the highest effectiveness is obtained. The prioritization is based on the combination of the importance and condition indexes. The combination is based on the ranking axe shown in Fig. 22. Perpendicular lines between the ordered pairs (II, CI) and the ranking axe are drawn and the resultant distance of the lines “d” is measured. As a result a vector D = [d1 , d2 , . . ., dn ]T is obtained (Eq. (8)). After sorting the elements of the vector D in descending order, the vector D is transformed into the vector Dranking = [dx , dy , . . .,dz ]T . The vector Dranking represents the ranking under which a prioritization of the maintenance actions (inspections, maintenance, replacements/repairs) have to be done. So for example, assuming that the vector Dranking corresponds to the transformers whose ordered pairs belong to the third zone, it could be concluded that the priority for the investments on replacement of transformers is as follows: first the transformer x, then the transformer y and as lasts the transformer z. The slope of the ranking axe represented by the angle ˇ determines the degree of influence of the importance and condition indexes for the prioritization of the maintenance actions. The angle ˇ varies in the interval from 0◦ till 90◦ . For ˇ = 0◦ , the influence of the importance index is disregarded what means that the prioritiza-

2034

J.L. Velasquez-Contreras et al. / Electric Power Systems Research 81 (2011) 2015–2037

tion of the activities are only carried out according to the condition index. In other words, ˇ = 0◦ corresponds to the condition-based maintenance strategy. In a similar way, for ˇ = 90◦ the influence of the condition index is disregarded. Considering that in the praxis both criteria (condition and importance) are expected to influence maintenance scheduling, as default a value for ˇ equals to 45◦ is recommended, since for this value of ˇ, the CI and the II have the same influence on the prioritization of activities. However, the degree of influence of the CI and II indexes can be set to particular preferences by means of the Eq. (8), where DI, with 0 ≤ DI ≤ 1, is the degree of influence that the importance index has on the prioritization. A DI = 0.5 means that the II has a 50% of influence on the prioritization and for this value of DI, ˇ = 45◦ . ˇ = 90 × DI

(8)

type (e.g. all 500 MVA transformers). These activities are closely related to one specific instance or one specific type of physical asset. Organisationally, they are usually under the responsibility of the maintenance department, but need some coordination with the operation department. Condition assessment and maintenance scheduling and execution are the main responsibilities of the maintenance department. • Network-oriented activities deal with the outage scheduling for the performance of maintenance and diagnostic tests, as well as with system constraints, such as maximum load and required network reliability. Organisationally, they are usually under the responsibility of the operations department, but need some coordination with maintenance. • Enterprise-oriented activities involve strategic decision making about investments, overall reliability and policies set up. They are rather managerial activities, and rely on the results of the other two groups.

8. GAMMEU in the environment of an electric utility This section presents how GAMMEU is adapted to the working context of an electrical utility. As indicated in [64], the activities related to maintenance in an electrical utility require the interaction of different organisational levels that are presented in Fig. 23. At the beginning of the useful life of the transformers, factory routine tests are performed by the manufacturer. Then, after the transportation, commissioning tests coordinated by the construction department are performed for assessing the integrity of the transformers and for generating the fingerprints or reference measurements under which the maintenance department will be able to assess further diagnostic measurements by means of comparisons to the fingerprints. After putting a transformer into service, its life cycle is managed by the departments of maintenance and operation, where the tasks carried out by these departments relies on the maintenance and operation policies of the utility (blocks 1 and 1 in Fig. 23). The maintenance activities can be gathered in three sectors, each of one has a specific orientation. These three sectors follow: • Asset-oriented activities focused on an asset as a component (e.g. transformer “xyz”), or on the population of assets of similar

A detailed description of the different elements shown in Fig. 23 is presented in [64]. For the purpose of this work, it is enough to understand that the maintenance activities are related with different organisational departments and the information flow among departments is of great relevance. With the aim of allowing GAMMEU to satisfy the different functions shown in Fig. 23, it was determined necessary to add the following three new elements within the model: • Enterprise Resource Planning System (EPRS) of the utility. This system is necessary for determining the availability of resources, the economical restrictions and for providing the information required for the cost estimation of the maintenance actions. • Utility Network Management System (NMS). This system consists of four major systems: Energy Management System (EMS), Transmission Management System (TMS), Distribution Management System (DMS) and Generation Management System (GMS). This system is in charge of carrying out the outage scheduling (blocks 2 and 4 in Fig. 23) as function of the operation policies. The term outage refers itself to taking the equipment out of service for the execution of maintenance works. Records of operations such as

Fig. 23. Maintenance related activities in an utility.

J.L. Velasquez-Contreras et al. / Electric Power Systems Research 81 (2011) 2015–2037

2035

Fig. 24. Asset manage model adapted to an electric utility environment.

disturbances or faults in the network are also managed by this system (block 3 ). • Utility Maintenance Management System (MMS). Through this system the work orders for the execution of the maintenance activities are generated and at the same time, the records of the different maintenance activities are generated.

ing approach for extracting the hidden knowledge in databases, which as before illustrated is the core for the development of intelligent systems for detection and diagnosis and failure rate estimation models.

After incorporating the three elements before described in GAMMEU, the resultant model is shown in Fig. 24, where the rows indicate the flow of data exchange among elements and the numbers inside the blocks associate the maintenance related activities shown in Fig. 24 to the elements of GAMMEU. The NMS and the MMS send the records of operation and maintenance to the platform for data integration respectively. The nameplate data of the transformers is to be directly uploaded in the platform for data integration from the NMS or from the MMS. Subsequently, the data are sent to the intelligent system for detection and diagnosis, where the condition indexes are calculated, which at the same time are sent to the failure rate estimation model for estimating the failure rate required for reliability analysis. The network data required for reliability analysis can be obtained from the NMS. The outcome of the reliability analysis is sent to the optimisation model for maintenance scheduling where the activities 1, 2, 4, and 5 are carried out. The scheduled maintenance actions are subsequently sent to NMS in order to schedule the required outages of the assets for execution of maintenance actions and/or diagnostic tests. Then maintenance tasks are scheduled by the MMS and work orders are generated for that purpose. The records of the executed maintenance actions are recorded by the MMS system itself and sent to platform of data integration for further condition assessment tasks. In summary, GAMMEU fulfils all the maintenance related activities shown in Fig. 24 and can be adapted to the typical information technology environment of utilities. The platform for data integration plays a fundamental role as the different components of GAMMEU are to be linked to existing IT systems in the utilities. The important data for the development of the elements of GAMMEU are the historical records of condition, maintenance and operation, which are normally stored in the databases of NMS and MMS systems. Application of artificial intelligence tools are a promis-

GAMMEU is a suitable framework for a successful implementation of asset management policies for power transformers in electric utilities due to its capability for considering and linking in a logical way all of the aspects related to the management of assets in a utility. The concept used for the design of the intelligent system based on multi-agents allows for a comprehensive detection and diagnosis of failure modes in the active part of power transformers. The detection task was carried out by the automatic and intelligent on-line monitoring system (AIOMS), which at the same time consists of a set of agents in charge of achieving specific detection tasks. The application of an artificial neural network trained using on-line measured data in a 30 MVA transformer illustrated the model-based monitoring concept for the detection of thermal failure modes. The validation of the model showed satisfactory results and the establishment of control limits and their interaction with an expert system under a multi-agent concept showed the capabilities of AIOMS. The design of the diagnostic module was carried out by means of agents for the assessment of both traditional and modern diagnostic methods. The assessment of the traditional tests is performed by the agents AIADGA, AIAPCE, AIAFDP, AIATET and AIAIRI for the assessments of dissolved gas analysis, physical chemical analysis, furan analysis and degree of polymerization, traditional electrical tests and infrared inspections, respectively. As knowledge bases for the development of these agents the expert knowledge collected by international standards as well as the hidden knowledge in databases of diagnostic tests is proposed as further research works. As modern diagnostic agents, AIAFRA, AIAFRS, AIADRM and AIAPD were formulated for the assessment of FRA, FRSL, DRM and PD measurements, respectively. The implementation of AIAFRA using decision trees for extracting the hidden knowledge in a data

9. Conclusions

2036

J.L. Velasquez-Contreras et al. / Electric Power Systems Research 81 (2011) 2015–2037

base of FRA measurements performed in real transformers was demonstrated. The obtained results are encouraging as the decision trees are able to perform an automatic assessment with a few classification rules. A novel methodology for characterizing the deterioration process of power transformers during its life span into different stages and states as a multi-state condition model (MSC model) was introduced. Taking as reference the MSC model and using the WCCM method introduced in this paper, a systematic combination of the individual diagnosis of diagnostic methods can be carried out. The application if this methodology was applied for a condition assessment of the degradation failure modes of a 30 MVA power transformer. As a result of the condition assessment, a condition index representing the overall condition of the transformer was obtained with a corresponding certainty factor. The proposed MSC model allows also training an algorithm based on a hidden Markov model under which the transition probabilities from one state to another can be determined and as a result, the time dependence of the failure rate of power transformers can be estimated. The implementation of the Markov model was exemplified using data of DGA tests performed in a 40 MVA transformer. The integration of all sub-systems of GAMMEU for supporting the decision making of the asset management of power transformers proposed in this paper relies on optimisation model that integrates all sub-systems, including reliability analysis. However, in view of the complexity behinds the practical implementation of the proposed model, a simple approach for maintenance scheduling using asset prioritization diagrams is also proposed as an alternative supporting the decision making. In summary, GAMMEU is a promising framework with built-in intelligence adapted to the actual necessities of the utilities with respect to reduction of maintenance costs and improvement of reliability. Implementations of some of its constituent elements in real transformers have demonstrated its technical feasibility and its capabilities. Further implementation of the whole elements of GAMMEU is object of further research, development and engineering works. References [1] J.L. Velásquez, et al., Guidelines for the implementation of condition monitoring systems in power transformers, in: Advanced Research Workshop on Transformers Baiona, Spain, 2007, pp. 265–270. [2] T. Orlowska, et al., Life cycle management of circuit-breakers by application of reliability centred maintenance, Cigré-Report 13-103, 2000. [3] Z. Wang, L. Tang, G. Frimpong, P. Tong Lee, Approaching decisions, Special report ABB review, 2003, pp.43–48. [4] IEEE/PES task force on impact of maintenance strategy on reliability, Impact of maintenance strategy on reliability, 1999. ˜ ADRES: a decision support system in the [5] F. Salamanca, I. González, A. Muina, updating of substations, Cigré-Report 23-105, 1996. [6] T. Kiiveri, M. Lahtinen, Life cycle cost and condition management system for substations, Cigré-Report 23-108, 1996. [7] J. McCalley, Y. Jiang, V. Honavar, J. Pathak, Automated Integration of Condition Monitoring with an Optimized Maintenance Scheduler for Circuit Breakers and Power Transformers, PSERC Publication 06-04, 2006. [8] M.C. Garcia, M.A. Sanz-Bobi, J. Del Pico, P. Sima, Intelligent System for Predictive Maintenance, Application to the health condition monitoring of a wind turbine, Comput. Ind. 57 (2006) 552–568. [9] Z. Zhang, Y. Jiang, J. MacCalley, Bayesian analysis of power transformers failure rate based on condition monitoring information, http://home.eng.iastate.edu/∼jdm/WebConfPapers/BayesianXfmrNAPS.pdf. [10] R.E.H. Brown, L. Willis, Failure rate modeling using equipment inspection data, IEEE Trans. Power Syst. 19–2 (2004) 782–787. [11] L. Ran, Ch. Li, G. Xueping, Condition-based maintenance of electrical equipments based on ID3 algorithm and Monte Carlo simulation, in: IEEE/PES Transmission and Distribution Conference & Exhibition: Asia and Pacific Dalian, China, 2005, pp. 1–5. [12] L. Bertling, R. Allan, R. Eriksson, A reliability-centered asset maintenance method for assessing the impact of maintenance in power distribution systems, IEEE Trans. Power Syst. 20-1 (2005) 75–82.

[13] J. Endrenyi, G.J. Anders, A.M. Leite da Silva, Probabilistic evaluation of the effect of maintenance on reliability—an application, IEEE Trans. Power Syst. 13-2 (1998) 576–583. [14] S. Vatland, et al., Utility systems integration, in: Nordic Distribution and Asset Management Conference, 2004, pp. 1–9. [15] J.A. Reinoso, A. Silvescu, D. Caragea, J.V. Pathak, G. Honavar, Information extraction and integration from heterogeneous, distributed autonomous information sources—a federated ontology-driven query-centric approach, in: IEEE International Conference on Information Reuse and Integration IRI, 2003, pp. 183–191. [16] M. Michalowski, et al., Retrieving and semantically integrating heterogeneous data from the web, IEEE Intell. Syst. 19-3 (2004) 72–79, doi:10.1109/MIS.2004.16. [17] J.A. Reinoso-Castillo, Ontology-driven information extraction and integration from heterogeneous distributed autonomous data sources: a federated query centric approach, Master thesis, Iowa State University, 2002. [18] S. Draber, E. Gelle, T. Kostic, O. Preiss, How operation data helps manage lifecycle costs, Cigré-Report 23/39-02, 2000. [19] J.J. Smit, Trends in emerging technologies in power systems, in: International Conference on Future Power Systems, 2005, pp. 1–7. [20] C. Liu, D.A. Pierce, H. Song, Intelligent system applications to power systems, IEEE Comput. Appl. Power 10-4 (1997) 22–24. [21] L. Wu, Y. Zhu, X. Li, J. Yuan, Application of multi-agent and data mining techniques in condition assessment of transformers, in: International Conference on Power System Technology—POWERCON, 2004. [22] Y. Zhu, L. Wu, X. Li, J. Yuan, A transformer condition assessment framework based on data mining, IEEE Power Engineering Society General Meeting (2005) 1875–1880. [23] C. Hung, M. Wang, Diagnosis of incipient faults in power transformers using CMAC neural network approach, Elec. Power Syst. Res. 71 (2004) 235– 244. [24] E. Mangina, Intelligent agent-based monitoring platform for applications in engineering, Int. J. Comput. Sci. Appl. 2-1 (2005) 38–48. [25] V.M. Catterson, S.E. Rudd, S.D.J. McArthur, G. Moss, On-line transformer condition monitoring through diagnostics and anomaly detection, in: The 15th International Conference on Intelligent System Applications to Power Systems, ISAP 2009, Curitiba, Brazil, November 8–12, 2009. [26] M. Wooldridge, An Introduction to Multiagent Systems, 2nd ed., Wiley, Great Britain, 2009. [27] R. Vilaithong, S. Tenbohlen, T. Stirl, Improved top-oil temperature model for unsteady-state conditions of power transformers, in: Proceedings of the XIVth International Symposium on High Voltage Engineering, Beijing, China, 2005, pp. 1–5. [28] B.C.W. Lesieutre, H.J. Hagman, L. Kirtley, An improved transformer top-oil temperature model for use in an online monitoring and diagnostic system, IEEE Trans. Power Deliv. 12-1 (1997) 249–256. [29] A. Ottele, R. Shoureshi, Neural network-based adaptive monitoring system for power transformer, Trans. ASME 123 (2001) 512–517. [30] V. Galdi, L. Lppolito, A. Piccolo, A. VAccaro, Neural diagnostic system for transformer thermal overload protection, IEE Proc. Electr. Power Appl. (2000) 147–155, doi:10.1049/ip-epa:20010086. [31] B. García, J.C. Burgos, A. Alonso, Winding deformations detection in power transformers by tank vibrations monitoring, Electr. Power Syst. Res. 74-1 (2005) 129–138. [32] Y. Kim, et al., Structure-vibration analysis of a power transformer (154 kV/60 MVA/single phase), in: Proceedings of World Academy of Science, Engineering and Technology, vol. 38, 2009, pp. 106–110. [33] T. Leibfried, K. Feser, Off-line and on-line-monitoring of power transformers using the transfer function method, in: Conference Record of the 1996 IEEE International Symposium on Electrical Insulation, Montreal, Canada, 1996, pp. 1333–1341, doi:10.1109/61.796226. [34] R. Wimmer, K. Feser, Calculation of the Transfer Function of a Power Transformer with Online Measuring Data, APTADM, Wroclaw, 2004. [35] IEEE guide for loading mineral-oil-immersed transformers, IEEE Std. C57.911995, 1995. [36] IEC 60354, Loading guide for oil-immersed power transformers, International Electrotechnical Commission Standard, Geneva, Suisse, 1991. [37] R. Pérez, E. Matos, S. Fernández, Identification of the nonlinear model proposed by the MIT for power transformers by applying genetic algorithms, IEEE Lat. Am. Trans. 7 (6) (2009) 636–642. [38] W.H. Tang, K. Spurgeon, Q.H. Wu, Z.J. Richardson, Modelling equivalent thermal dynamics of power transformers using genetic algorithms, in: International Conference on Power System Technology, 2002. Proceedings. PowerCon 2002, vol. 3, 2002, pp. 1396–1400, doi:10.1109/ICPST.2002.1067759. [39] A. Ottele, R. Shoureshi, Neural network-based adaptive monitoring system for power transformer, Trans. ASME 123 (September) (2001) 512–517. [40] T.C.B.N. Assunc¸ão, J.L. Silvino, P. Resende, Transformer top-oil temperature modeling and simulation, Trans. Eng. Comput. Technol. 15 (2006) 240–245. [41] M.K. Pradhan, T.S. Ram, On-line monitoring of temperature in power transformers using optimal linear combination of ANNs, in: Conference Record of the 2004 EFE International Symposium on Electrical Insulation, Indianapolis, USA, September 19–22, 2004, pp. 70–73. [42] R. Vilaithong, S. Tenbohlen, T. Stirl, Neural network for transformer top-oil temperature prediction, in: XVth International Symposium on High Voltage Engineering, University of Ljubljana, Elektroinˇstitut Milan Vidmar, Ljubljana, Slovenia, August 27–31, 2007, pp. 1–6.

J.L. Velasquez-Contreras et al. / Electric Power Systems Research 81 (2011) 2015–2037 [43] J.L. Velásquez, et al., Development and implementation of a condition monitoring system in a substation, in: 9th International Conference on Electrical Power Quality and Utilization, Barcelona, Spain, 2007. [44] M. Hässig, et al., On-site applications of advanced diagnosis methods for quality assessment of insulation of power transformers, in: Conference on Electrical Insulation and Dielectric Phenomena, 2001, pp. 441–447, doi:10.1109/CEIDP.2001.963577. [45] M. Duval, A. dePablo, Interpretation of gas-in-oil analysis using new IEC publication 60599 and IEC TC 10 databases, IEEE Electr. Insul. Mag. 17-2 (2001) 31–41. [46] IEEE Std 57.106TM -2002, IEEE guide for acceptance and maintenance of insulating oil in equipment, 2002. [47] IEEE Std 62TM -1995, IEEE guide for diagnostic field testing of electric power apparatus—part 1: oil filled power transformers, regulators, and reactors, 1995. [48] J.L. Velásquez, M.A. Sanz-Bobi, M. Gutierrez, A. Kraetge, Knowledge bases for the interpretation of the frequency response analysis of power transformers, in: Congreso Internacional en alta tensión y aislamiento eléctrico, ALTAE 2009, Medellín, Colombia, November 23–27, 2009. [49] J. L. Velásquez, D. Kolb, M.A, Sanz-Bobi, W. Koltunowicz, Sub-band identification of transformer-specific frequency sub-bands as basis for a reliable and automatic assessment of FRA results, Presented in the CDM Conference, CDM 2010, Tokio, Japan, September 6–11, 2010. [50] J.R. Quinlan, C4. S 5: Programs for Machine Learning, Morgan Kaufmann Publishers Inc., USA, 1993. [51] C.M. Bishop, Pattern Recognition and Machine Learning (Information Science and Statistics), Springer, USA, 2006. [52] D. Morais, J. Rolim, Z. Vale, DITRANS—a multi-agent system for integrated diagnosis of power transformers, 1–6, in: IEEE Powertech (POWERTECH 2008), vol. 5, Lausanne, Switzerland, 2008. [53] Hydro plant risk assessment guide, Appendix E5: transformer condition assessment, 2006. http://www.docstoc.com/docs/7274124/Hydro-Plant-RiskAssessment-Guide-Appendix-E6-Turbine-Condition. [54] N. Dominelli, A. Rao, P. Kundur, Life extension and condition assessment, IEEE Power Energy Mag. 25 (2006) 25–35. [55] CIGRÉ Working group A2.18, Guide for life management techniques for power transformers, 2003. [56] P. Loring, Extend the lifetime of your transformers by using computermodelling, IEE Power Eng. 19-2 (2005) 18–21, doi:10.1049/pej:20050203. [57] J. Pathak, Y. Jiang, V. Honavar, J. McCalley, Condition data aggregation with application to failure rate calculation of power transformers, in: Proceedings of the 39th Hawaii International Conference on System Sciences, 2006, pp. 1–10. [58] F. Sotiropoulos, P. Alefragis, E. Housos, A Hidden Markov models tool for estimating the deterioration level of a power transformer, in: IEEE Confer-

[59] [60] [61] [62] [63] [64]

2037

ence on Emerging Technologies and Factory Automation, 2007, pp. 784–787, doi:10.1109/EFTA.2007.4416857. J.M.H. Bollen, Literature search for reliability data of components in electric distribution networks, EUT Report 93 E-276, 1993. ISBN 90-6144-276-1. A review of the reliability of electric distribution system components: EPRI white paper, EPRI, Palo Alto, CA, 2001. 1001873. IEEE Std 493-1990, IEEE recommended practices for design of reliable industrial and commercial power systems, 1990. J.B. Bowles, Commentary-caution:, Constant failure-rate models may be hazardous to your design, IEEE Trans. Rel. 51 (2002) 375–377. P. Hilber, Maintenance optimization for power distribution systems, Doctoral thesis, Royal Institute of Technology, Stockholm, Sweden, 2008. T. Kostic, Asset management in electrical utilities: how many facets it actually has, in: Power Engineering Society General Meeting, 2003, pp. 275–281.

Juan L. Velásquez Contreras was born in Venezuela in 1980. He received the M.Sc. degree in Electrical Engineering from the UNEXPO in 2002. Afterwards he joined CVG Venalum in Venezuela, where he worked as maintenance and project engineer of high voltage assets. In 2006 he joined CITCEA, centre of technological innovation in Spain, where he worked as project engineer for the implementation of condition monitoring systems in power transformers. Since October 2008 till Octuber 2011 he worked by Omicron Electronics in Austria as product manager of diagnostic instruments. He is currently working towards the Ph.D. degree in the area of Monitoring and Diagnosis of Power Transformers at the Polytechnic University of Catalonia, in Barcelona, Spain. Miguel A. Sanz Bobi is a professor at the Computer Science Department and also researcher at the Institute for Research and Technology (IIT) both inside the Engineering School of the Pontificia Comillas University, Madrid (Spain). He divides his time between teaching and research in the artificial intelligence field applied to diagnosis and maintenance of industrial processes. He has been the main researcher in more than 35 industrial projects over the last 20 years related to the diagnosis in real-time of industrial processes, incipient detection of anomalies based on models, knowledge acquisition and representation, reliability and predictive maintenance. All these projects have been based on a combination of artificial intelligence, new information technologies and data mining techniques. Samuel Galceran Arellano was born in Lleida, Spain, in 1971. He received the M.Sc. degree in electrical engineering and the Ph.D. degree from the Polytechnic University of Catalonia (UPC), Barcelona, Spain, in 1997 and 2002, respectively. In 1997, he joined the Electrical Engineering Department, UPC, as an Assistant Professor. He developed several projects for industry, and in 2001, he joined the Center of Technological Innovation in Static Converters and Drives (CITCEA), UPC, where he belongs to the CITCEA Directorate staff. His primary research interests are motor control and converters for power supplies and drives.