Produção Nacional
Revisado por pares
Islanding detection of synchronous distributed generators using data mining complex correlations
2018; Institution of Engineering and Technology; Volume: 12; Issue: 17 Linguagem: Inglês
10.1049/iet-gtd.2017.1722
ISSN1751-8695
AutoresEduardo A. P. Gomes, José Carlos de Melo Vieira Júnior, Denis V. Coury, Alexandre C. B. Delbem,
Tópico(s)Islanding Detection in Power Systems
ResumoIET Generation, Transmission & DistributionVolume 12, Issue 17 p. 3935-3942 Research ArticleFree Access Islanding detection of synchronous distributed generators using data mining complex correlations Eduardo A.P. Gomes, Corresponding Author Eduardo A.P. Gomes eduardoapgomes@usp.br Department of Electrical and Computer Engineering, São Carlos School of Engineering, University of São Paulo, Av. Trabalhador São-Carlense, 400, 13566-590 São Carlos, SP, BrazilSearch for more papers by this authorJosé C.M. Vieira, José C.M. Vieira Department of Electrical and Computer Engineering, São Carlos School of Engineering, University of São Paulo, Av. Trabalhador São-Carlense, 400, 13566-590 São Carlos, SP, BrazilSearch for more papers by this authorDenis V. Coury, Denis V. Coury Department of Electrical and Computer Engineering, São Carlos School of Engineering, University of São Paulo, Av. Trabalhador São-Carlense, 400, 13566-590 São Carlos, SP, BrazilSearch for more papers by this authorAlexandre C.B. Delbem, Alexandre C.B. Delbem Institute of Mathematical and Computer Sciences, University of São Paulo, Av. Trabalhador São-Carlense, 400, 13566-590 São Carlos, SP, BrazilSearch for more papers by this author Eduardo A.P. Gomes, Corresponding Author Eduardo A.P. Gomes eduardoapgomes@usp.br Department of Electrical and Computer Engineering, São Carlos School of Engineering, University of São Paulo, Av. Trabalhador São-Carlense, 400, 13566-590 São Carlos, SP, BrazilSearch for more papers by this authorJosé C.M. Vieira, José C.M. Vieira Department of Electrical and Computer Engineering, São Carlos School of Engineering, University of São Paulo, Av. Trabalhador São-Carlense, 400, 13566-590 São Carlos, SP, BrazilSearch for more papers by this authorDenis V. Coury, Denis V. Coury Department of Electrical and Computer Engineering, São Carlos School of Engineering, University of São Paulo, Av. Trabalhador São-Carlense, 400, 13566-590 São Carlos, SP, BrazilSearch for more papers by this authorAlexandre C.B. Delbem, Alexandre C.B. Delbem Institute of Mathematical and Computer Sciences, University of São Paulo, Av. Trabalhador São-Carlense, 400, 13566-590 São Carlos, SP, BrazilSearch for more papers by this author First published: 24 May 2018 https://doi.org/10.1049/iet-gtd.2017.1722Citations: 11AboutSectionsPDF ToolsRequest permissionExport citationAdd to favoritesTrack citation ShareShare Give accessShare full text accessShare full-text accessPlease review our Terms and Conditions of Use and check box below to share full-text version of article.I have read and accept the Wiley Online Library Terms and Conditions of UseShareable LinkUse the link below to share a full-text version of this article with your friends and colleagues. Learn more.Copy URL Share a linkShare onFacebookTwitterLinkedInRedditWechat Abstract This study proposes a novel method based on data mining for islanding detection of synchronous distributed generators. The method uses a new technique called data mining of code repositories (DAMICORE), which is a powerful data mining tool for detecting patterns and similarities in various kinds of datasets. In addition, a trip logic was developed in this proposal in order to detect islanding and disconnect the distributed generator. One of the most relevant features of the proposed method is its capability to generalise, which reduces the need of big datasets for training purposes. This approach has been tested with islanding, load switching and fault simulations, presenting promising results concerning performance, as well as detection time. General results showed a better performance of the method if compared to traditional anti-islanding protection schemes, such as frequency-based relays. 1 Introduction Modern distribution systems need major initiatives related to the concept of smart grids. Some of the challenges faced by this new reality are: managing an active distribution power grid with telecommunication infrastructure, the existence of distributed resources, as well as the increasing number of customers demanding high power quality. In this context, distributed generation (DG), energy storage, electric vehicles, demand response and smart meters play an important role in this new smart grid scenario. These technologies affect both the planning functions and real-time operation of the system [1–3]. DG has a significant impact on the development of smart grids. As an example, electrical losses can be reduced and investments in transmission lines and centralised generation can be postponed. However, DG faces voltage regulation and protection issues [4]. Thus, distributed generators should be installed based on in-depth studies in order to identify the impact on the distribution grid operation. An important DG issue is to protect distributed generators against islanding. The IEEE 1547 [5] standard defines islanding as a condition in which a portion of the power distribution system becomes energised solely by one or more distributed generators, making an energised island disconnected from the main supply. Inadvertent islanding causes some problems concerning safety, commercial issues, power quality and system integrity. In summary, the main issues include the maintenance teams and public safety, miscoordination of the overcurrent protection devices, inadequate grounding, as well as out-of-phase reclosing [6]. Generally, islanding detection is performed by passive protection, such as frequency, the rate of change of frequency, vector surge and voltage protection [7]. These schemes tend to fail if the power imbalance between the islanded load and the distributed generator is small. These failures can be minimised if the protection settings are adjusted to an over sensitive level, but this can cause nuisance tripping. Therefore, there must be a trade-off between anti-islanding protection sensitivity and selectivity, which is hard to achieve by using the traditional frequency and voltage-based anti-islanding protection. On the other hand, the passive anti-islanding detection techniques are more economically feasible than the remote ones, and they do not insert voltage or current disturbances in the power distribution system, as the active techniques do [8]. Recently, new types of passive anti-islanding techniques based on data mining and artificial intelligence have emerged as potential solutions to detect islanding efficiently. Those approaches claimed to be immune to transients caused by events other than islanding. The data mining based support vector machine (SVM), decision tree (DT) and random forest (RF) methods were tested in [9], in which DT has shown the advantage of being an if-else based rules solution in the hardware environment. Meanwhile, SVM and RF methods provided a similar performance compared to the DT method. The DT in [10] faces a challenge to find the optimal configuration of the if-else based rules, which are able to classify islanding properly. Generally, DT-based methods perform correct classification considering specific cases within the system in which they were trained. However, if the topology or the operation condition of the system changes, then the classification performance of the DT can be reduced. Dash et al. [11] use the DT results to create a fuzzy-based islanding detection scheme, which leads to a better performance solution than the stand-alone DT. In [12], several DTs are used to create an adaptive anti-islanding protection scheme capable of self-determining the trip logic and its threshold as a function of the power imbalance in the islanding. However, SVM, DT and RF have some drawbacks. Some examples are as follows: SVM training is not practical if the problem needs a large amount of data; DT is susceptible to overfitting and requires large databases for training and validation procedures; RF shares the same issues as DT because it combines several DTs [13, 14]. Artificial neural network (ANN) based techniques are presented in [15, 16]. The ANN proposed by Merlin et al. [15] uses the voltage waveform with windows of 64 and 128 samples per cycles measured at the distributed generator point of common coupling. In [16], the optimisation techniques (evolutionary programming, particle swarm optimisation) are used to find efficient settings of the proposed ANN. Both methods [15, 16] deliver a good performance concerning islanding detection. However, ANN-based methods demand costly training adjustments and optimisation procedures to find the best settings of the neural network. Other intelligent islanding detection schemes using preprocessing methods such as wavelet transform (WT), S-transform (ST), time-time transform (TTT), mathematical morphology and autoregressive (AR) coefficients must be mentioned. Such methods demand a high computational hardware performance and are more sensitive to noise effects and harmonic distortion. Ray et al. [17] proposed a WT and an ST based approaches for islanding detection. In both cases islanding can be detected by using the voltage waveform. However, some drawbacks of this approach are: the performance of WT is influenced by noise and the ST demands more computational power than WT. Two WT-based islanding detection approaches are proposed by Alshareef et al. [18]. The authors tested the SVM and DT classification methods combined with the WT voltage preprocessing. In this case, the SVM-based approach presented superior performance than the DT-based method. Mohanty et al. [19] compared many advanced preprocessing techniques combined with SVM in order to classify islanding and other power quality disturbances. The preprocessing techniques WT, ST, HST, TTT and mathematical morphology were used in the voltage waveform. TTT and mathematical morphology were more efficient to detect islanding than HST and ST. The drawbacks of such preprocessing techniques are: both performances of WT and ST are influenced by noise. Harmonics also have an impact on the performance of WT. The HST, TTT and mathematical morphology methods are less sensitive to noise and harmonics than WT and ST. Matic-Cuka and Kezunovic [20] combined the SVM and AR coefficients of voltage and current signals to design a new anti-islanding detection method for single-phase inverter based technologies. The main drawback is that the performance of the AR coefficients is influenced by the sliding window size. However, such an approach detects islanding properly considering low reactive and active power mismatches. In this paper, data mining of code repositories (DAMICORE) is investigated as a potential tool to detect islanding of distributed synchronous generators (SG). DAMICORE is a powerful unsupervised classification tool capable of detecting similarities in various kinds of datasets. In [21, 22], it is shown that the method has been successfully applied to complex datasets in problems related to computer science. In the context of islanding detection, DAMICORE is able to separate islanding from non-islanding situations, by processing the data correlation of the signal samples used. This paper also presents principles and tests of a local and passive islanding detection protection scheme based on DAMICORE. As main contribution, this paper proposes a new anti-islanding DAMICORE based detection method. The major advantages of the method are local preprocessing, the reduced number of simulations needed for training purposes and its high generalisation capacity. The method also provides a graphical analysis tool capable of showing the relationship among multivariable type data. The tests carried out to assess the effectiveness of the DAMICORE-based anti-islanding protection include more than a thousand simulations (including islanding, load switching and faults). The analysis of the DAMICORE-based protection dependability, safety and accuracy was also considered. Moreover, its performance was compared with frequency-based relays commonly used in DG anti-islanding protection, such as standard frequency relays. The results showed that the proposed scheme has a high potential for detecting DG islanding, presenting a better performance than the traditional methods in terms of both speed and dependability. This paper is organised as follows. Section 2.1 describes the process of Knowledge Discovery in Databases (KDD). Section 2.2 introduces the data mining technique used. Section 3 proposes a DAMICORE-based KDD for islanding detection. Section 4 presents the results and Section 5 draws the conclusions. 2 Basic concepts relating KDD and DAMICORE Data mining techniques refer to a particular step of the KDD. The KDD is a process that summarises the steps necessary for applications of data mining algorithms in order to extract useful knowledge from any database. Fig. 1a introduces the steps of KDD, where DAMICORE has been chosen as the data mining technique. The results from the DAMICORE-based KDD process are then used to formulate an intelligent islanding detection scheme. Fig. 1Open in figure viewerPowerPoint DAMICORE-based KDD process for the anti-islanding protection scheme (a) KDD process [23], (b) DAMICORE algorithm 2.1 KDD process The following steps describe the KDD process [23]: Selection: It creates the target dataset by selecting specific data samples from a database. Preprocessing: It performs the feature extraction of data samples. Furthermore, it addresses the noise, redundancy or incomplete data samples and removes the outliers from the target dataset. Transformation: It arranges the preprocessed target dataset to compose the inputs for the data mining algorithm. For example, feature selection methods can be used. Data mining: It analyses the transformed data in order to find patterns that can benefit the KDD process. Interpretation: An expert analyses the mined patterns aiming to develop an algorithm that generates automatic analysis reports or automatic decisions. 2.2 Data mining of code repositories DAMICORE is a data mining technique that requires no training or previous knowledge in order to find relevant relationships in data. Fig. 1b summarises the DAMICORE steps. At first, DAMICORE requires the data to be organised in files inside a folder. Then, three essential steps carry out the discovery of patterns. The first step estimates the similarity level of each pair of files using a dissimilarity metric, such as the Euclidean distance (ED), producing a distance matrix. Then, neighbour joining (NJ) algorithm [24] constructs a phylogenetic tree from the distance matrix. In the field of phylogeny, the phylogenetic tree is a graphical representation of the relationships among living beings with a common ancestor [24]. In DAMICORE, the phylogenetic tree synthesises the levels of correlation among the files. The fast Newman (FN) algorithm [25] uses the phylogenetic tree graph for clustering. Finally, DAMICORE produces two outputs: the first one is a set of files organised in clusters and the second one presents the phylogenetic tree highlighting the possible relations among the clusters found. The patterns found by DAMICORE can benefit the learning process from data and the generalisation capacity of prediction models based on it. Section 3 presents an approach based on DAMICORE. 3 Local data mining anti-islanding system for SG (LDMAIS-SG) The LDMAIS-SG proposed in Fig. 2 has offline and online steps, as explained below: Offline steps: They comprise the training steps by using DAMICORE to detect patterns of islanding, non-islanding and suspect islanding. The voltage and current signals from selected simulations are preprocessed and stored in order to generate input text files used by DAMICORE. The patterns found by the DAMICORE are stored in a hardware memory of the data mining function, enabling LDMAIS-SG to operate online. The offline steps require the KDD process to be performed in sequence: selection (Section 3.1); transformation (Section 3.3); DAMICORE (Section 2.2) and interpretation (Section 3.4); Online operation: Each new sample of voltages and currents from the current transformers and potential transformers are preprocessed locally, transformed and stored in a circular buffer, i.e. each new sample stored in the buffer discards the old one. Thus, the generation of input files is no longer necessary in the online operation because all the data used for the data mining function have been obtained in the offline step and are stored in the memory. Finally, the samples stored in the circular buffer are used by the data mining function to detect islanding. The online operation of LDMAIS-SG performs the algorithm's preprocessing (Section 3.2), transformation (Section 3.3) and data mining function (Section 3.5) in sequence. Fig. 2Open in figure viewerPowerPoint Local data mining anti-islanding system for SG 3.1 Selection The first step of the off-line KDD process is to select specific simulations of islanding and other disturbances and store the three phase voltage and current signals measured at the DG location. This step is important because the method learns how to cluster new samples based on the selected disturbances. The more distinct the disturbances are, the more general the clustering results (produced by DAMICORE) are. Thus, for the LDMAIS-SG to perform the islanding detection in other distribution power systems (DPS), simulations of load switching, faults, and different cases of islanding are necessary for the selection step. 3.2 Preprocessing The locally measured voltages and currents need to be processed in order to create the feature vector. Fig. 3a shows how this vector is built. The voltage and current waveforms are sampled at . The up and down zero crossing detection block is used to trigger the buffers so that they store one cycle of the voltage and current waveforms. Afterwards, the following features are computed by using the information of one cycle and they are updated at each half cycle: x0 : f = frequency (Hz); x1 : (df/dt) = rate of change of x0 (Hz/s); x2 : V = RMS voltage (p.u.); x3 : (dV/dt) = rate of change of x2 (p.u/s); x4 : P = active power (p.u.); x5 : Q = reactive power (p.u).; x6 : cos(θ) = power factor; x7 : (dPF/dt) = rate of change of x6 (1/s); x8 : Δθ = vector surge (deg); x9 : (dΔθ/dt) = rate of change of x8 (1/s). These ten features were selected because they are representative quantities that can define an islanding state. To remove noise, two low-pass Butterworth filters are used: the input filter is a second order with a cutoff frequency of 120 Hz and the output filter is a first order with a cutoff frequency of 10 Hz. Fig. 3Open in figure viewerPowerPoint Preprocessing and transformation algorithm (a) Feature vector algorithm, (b) Transformation algorithm 3.3 Transformation The input files for DAMICORE and the data mining function are available by running the transformation algorithm of Fig. 3b, considering all the preprocessed simulation data in order to generate the feature matrices. Each feature matrix contains n lines (cycles of feature vector) and ten columns (one per feature). The first step specifies the quantity of cycles and normalises the features to produce more consistent signal comparisons with the ED. The normalisation procedure divides each feature by the correspondent maximum value in the simulations of the selection step (see Section 3.1). Then, the transformation algorithm constructs a feature matrix by concatenating the feature vector during n cycles. The final step stores the feature matrix. In the offline steps, the feature matrix is stored in an input file (in text format), while in the online operation it is stored in a circular buffer and in the memory of the data mining function. 3.4 Interpretation For each input file, the interpretation process performs the labelling of the mined patterns found by DAMICORE into three classes (islanding, suspect islanding and non-islanding), as follows: Islanding cases (IL): If the mined patterns contain only feature matrices of islanding; Suspect islanding cases (SIL): If the mined patterns contain feature matrices from both islanding and other disturbances. SIL covers cases in which DAMICORE confuses the classification of islanding with non-islanding events and, therefore, they are grouped; Non-islanding cases (NIL): If the mined patterns contain only feature matrices from disturbances different from islanding. Afterwards, the resulting knowledge of the classes concerning the database is stored in the data mining function and the KDD process with DAMICORE stops running. 3.5 Data mining islanding detection function Fig. 4 presents the online operation of the data mining function. The algorithm decides if it is islanding or not by checking all the EDs and the corresponding classes between the input feature matrix in the circular buffer and the feature matrices interpreted by the DAMICORE-based KDD process in the memory. Then, the algorithm performs the classification of the input feature matrix into IL, SIL or NIL by using the corresponding class of the minimum value among the EDs (MinED). Two adjustments are necessary for the algorithm to detect islanding safely as fast as possible. They are the MinED threshold (MinDist) and the SIL counter threshold (TripAlert), as explained in Sections 3.5.1 and 3.5.2, respectively. Thus, there are two if-based decisions for the algorithm to detect islanding: if the input feature matrix belongs to the islanding class and the correspondent minimum ED does not exceed the MinDist threshold, then islanding is detected (TRIP), or, if the SIL counter (TripCounter) exceeds the TripAlert threshold, then islanding is detected (TRIP). Moreover, suspect islanding could be one of these situations: if the input feature matrix belongs to SIL, then increase the TripCounter, or, if the input feature matrix belongs to the IL class and the correspondent minimum ED exceeds the MinDist threshold, then increase the TripCounter. The process of adjusting MinDist and TripAlert consists of analysing islanding detection concerning the simulation cases defined in the selection step (Section 3.1). If the data mining function detects islanding correctly for all the situations, the setting procedures are complete. Sections 3.5.1 and 3.5.2 introduce the adjustment procedures of MinDist and TripAlert thresholds. Fig. 4Open in figure viewerPowerPoint Data mining islanding detection function (data mining function) 3.5.1 MinDist adjustment Accurate islanding classification occurs when the input feature matrix belongs to the islanding class and the respective minimum ED is equal to zero. In this condition, the input feature matrix (online) stored in the circular buffer and at least one of the feature matrices (offline) stored in the memory are exactly the same. However, small time delays may occur between the online and offline feature matrices during transients, which causes . The zero-crossing based preprocessing technique was designed to compensate this delay, but does not eliminate it. By observing the MinDist behaviour considering the online operation with the same simulations of the KDD process, the limit found for MinDist was <3. Thus, the value of 3 is a dimensionless result of the ED and it is considered as an acceptable time delay difference compensation in this application. Once the feature matrices are normalised, a MinDist equal to 3 instead of zero can be used for testing the proposed technique in other distribution systems. 3.5.2 TripAlert adjustment The TripAlert is adjusted for the algorithm to detect islanding in situations where it is very similar to other disturbances. The data mining function detects islanding if a suspect islanding situation persists for a certain period of time. Two data mining functions are designed. The first operates with five cycles and the second with ten cycles. For each one, a different value of TripAlert was specified. In the 5-cycle data mining function, the TripAlert is 3, while in the 10-cycle data mining function, the TripAlert is 2. The TripAlert can be adjusted by checking if islanding is detected correctly for all the training simulations. 4 Results The LDMAIS-SG (proposed anti-islanding protection) is tested in a DPS modelled in SimPowerSystems using MATLAB R2010a. The single line diagram in Fig. 5 represents a sub-transmission power system, with a nominal voltage of 138 kV, frequency 60 Hz, short circuit level 500 MVA that feeds a DPS where a distributed SG of 12.6 MVA is installed at bus B4. The SG is equipped with an automatic voltage regulator and IEEE DC1A excitation system [26]. The tandem-compound (single mass) primary mover system is represented by a speed regulator, steam turbine and a shaft [27]. The distribution lines have an impedance of and other system data can be found in the Appendix. Fig. 5Open in figure viewerPowerPoint Single line diagram of the power system Table 1 presents the quantity and type of simulations considered in the KDD selection step for the DPS in this paper. To run the 12 islanding cases, the DG power and/or the load power were varied. This has also been done to simulate the non-islanding events. A two-second data window was used to feed the DAMICORE algorithms: 1 s before the disturbance (steady-state condition) and after the disturbance. Table 1. Number of simulations for training Event Number of imulations islanding 12 load switch opening 6 load switch closing 6 fault 15 total 39 The following sections present the analysis of the DAMICORE data mining results and subsequently the protection tests. 4.1 Data mining analysis The input files for the data mining process with DAMICORE follow the offline steps of KDD presented in Section 2.1: selection, preprocessing, transformation and data mining. Thus, this section concerns to show the results of the interpretation step of the data mining results. Two tests were run to investigate the influence of the feature vector number of cycles in the DAMICORE capability to identify islanding situations. The first test considered n equal to 5 cycles and the second considered n equal to 10 cycles. Both results can be observed in Fig. 6. This figure shows two phylogenetic trees and each one represents the input data, which were organised in files. These files represent the simulations of Table 1 and steady-state conditions. In the leaves of the trees, there are coloured identifications of the files (their names), representing the classes of the simulated events, as presented below: Red: islanding Gold: phase-to-ground fault Dark blue: load switch opening Light blue: load switch closing Green: steady-state The phylogenetic trees are adequate graphical tools to represent the correlation of a large amount of data. Their internal nodes account for the data correlation. A clade is a set of leaf nodes in the tree grouped together by one internal node. If a clade has many leaves, it means that the files (data) have high correlation levels, so that they can be grouped into a cluster. Thus, an easy way to visually interpret those trees is to analyse the sets of files of the same colour that are grouped. Fig. 6Open in figure viewerPowerPoint Interpretation of DAMICORE results by phylogenetic trees (a) Feature matrices: 5 × 10, (b) Feature matrices: 10 × 10 In Fig. 6a, there is a large set of clades with many leaves classified as islanding and there is a clade with a few number of leaves, also classified as islanding, but not grouped together with the large set of islanding clades. This indicates that the method did not classify all the islanding situations correctly. However, notice that as the number of cycles increases from 5 to 10 cycles (Fig. 6b), this error tends to vanish. This occurs because the information in one post-disturbance cycle is not enough to characterise an event confidently, i.e. the first cycle of an islanding situation can be easily confused with one cycle of load switching or fault situations. This confusion tends to diminish as the number of cycles increases. However, the larger the number of cycles, the slower the method takes to detect an islanding condition. Thus, the choice of the number of cycles is a trade-off between accuracy and speed. The final results of DAMICORE are the mined patterns found in the phylogenetic tree by the FN technique. Table 2 presents the patterns that can be interpreted in the following classes: IL, SIL or NIL as defined previously. DAMICORE found nine patterns on both trees in Fig. 6. Three patterns contain groups of single classes, including islanding, load switching and phase-to-ground faults. Four patterns contain double class data (4, 5, 6 and 7). Two patterns (8 and 9) contain triple class data. Pattern eight contains cases of islanding and can be found only in Fig. 6a. The DAMICORE did not find pattern eight in Fig. 6b, confirming that the classification performance is better considering more cycles. Table 2. Cluster characteristics of DAMICORE Single classes Double classes Triple classes 1 2 3 4 5 6 7 8 9 IL F LS IL IL F SS IL F — — F LS LS LS F RP — — — — — — LS LS ab ab ab ab ab ab ab a ab IL, islanding; F, phase-to-ground fault; LS, load switching; SS, steady state; a, tree a ; b, tree b ; bold, islanding; italic, suspect islanding; non-bold, non-italic, non-islanding. 4.2 General performance tests In this section, the islanding detection performance of the proposed method is compared with the conventional frequency-based relays. For the proposed data-mining scheme, 5 and 10 cycles were used. The frequency based relays were modelled in MATLAB by using the frequency outpu
Referência(s)